Академический Документы
Профессиональный Документы
Культура Документы
Application to
Differential Equations
Copyright © 2008 by Morgan & Claypool
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means—electronic, mechanical, photocopy, recording, or any other except for brief quotations in
printed reviews, without the prior permission of the publisher.
DOI 10.2200/S00146ED1V01Y200808MAS002
Lecture #2
Series Editor: Steven G. Krantz, Washington University, St. Louis
Series ISSN
Synthesis Lectures on Mathematics and Statistics
ISSN pending.
Jordan Canonical Form:
Application to
Differential Equations
Steven H. Weintraub
Lehigh University
M
&C Morgan & cLaypool publishers
ABSTRACT
Jordan Canonical Form ( JCF) is one of the most important, and useful, concepts in linear algebra.
In this book we develop JCF and show how to apply it to solving systems of differential equations.
We first develop JCF, including the concepts involved in it–eigenvalues, eigenvectors, and chains
of generalized eigenvectors. We begin with the diagonalizable case and then proceed to the general
case, but we do not present a complete proof. Indeed, our interest here is not in JCF per se, but in
one of its important applications. We devote the bulk of our attention in this book to showing how
to apply JCF to solve systems of constant-coefficient first order differential equations, where it is
a very effective tool. We cover all situations–homogeneous and inhomogeneous systems; real and
complex eigenvalues. We also treat the closely related topic of the matrix exponential. Our discussion
is mostly confined to the 2-by-2 and 3-by-3 cases, and we present a wealth of examples that illustrate
all the possibilities in these cases (and of course, a wealth of exercises for the reader).
KEYWORDS
Jordan Canonical Form, linear algebra, differential equations, eigenvalues, eigenvectors,
generalized eigenvectors, matrix exponential
v
Contents
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .v
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
A Background Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
A.1 Bases, Coordinates, and Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
A.2 Properties of the Complex Exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .85
vi CONTENTS
Preface
Jordan Canonical Form ( JCF) is one of the most important, and useful, concepts in linear
algebra. In this book, we develop JCF and show how to apply it to solving systems of differential
equations.
In Chapter 1, we develop JCF. We do not prove the existence of JCF in general, but
we present the ideas that go into it—eigenvalues and (chains of generalized) eigenvectors. In
Section 1.1, we treat the diagonalizable case, and in Section 1.2, we treat the general case. We
develop all possibilities for 2-by-2 and 3-by-3 matrices, and illustrate these by examples.
In Chapter 2, we apply JCF. We show how to use JCF to solve systems Y = AY + G(x) of
constant-coefficient first-order linear differential equations. In Section 2.1, we consider homoge-
neous systems Y = AY . In Section 2.2, we consider homogeneous systems when the characteristic
polynomial of A has complex roots (in which case an additional step is necessary). In Section 2.3,
we consider inhomogeneous systems Y = AY + G(x) with G(x) nonzero. In Section 2.4, we
develop the matrix exponential eAx and relate it to solutions of these systems. Also in this chapter
we provide examples that illustrate all the possibilities in the 2-by-2 and 3-by-3 cases.
Appendix A has background material. Section A.1 gives background on coordinates for
vectors and matrices for linear transformations. Section A.2 derives the basic properties of the
complex exponential function. This material is relegated to the Appendix so that readers who
are unfamiliar with these notions, or who are willing to take them on faith, can skip it and still
understand the material in Chapters 1 and 2.
Our numbering system for results is fairly standard: Theorem 2.1, for example, is the first
Theorem found in Section 2 of Chapter 1.
Steven H. Weintraub
Lehigh University
Bethlehem, PA USA
July 2008
viii PREFACE
1
CHAPTER 1
Av = λv ,
5 −7 7
Example 1.2. Let A be the matrix A = . Then, as you can check, if v1 = , then
2 −4 2
1
Av1 = 3v1 , so v1 is an eigenvector of A with associated eigenvalue 3, and if v2 = , then
1
Av2 = −2v2 , so v2 is an eigenvector of A with associated eigenvalue −2.
Av = λv
Av = λI v
(A − λI )v = 0 .
Eλ = {v | Av = λv} = {v | (A − λI )v = 0} = Ker(A − λI ) .
Definition 1.3. The characteristic polynomial of a matrix A is the polynomial det(λI − A).
2 CHAPTER 1. JORDAN CANONICAL FORM
Remark 1.4. This is the customary definition of the characteristic polynomial. But note that, if A is
an n-by-n matrix, then the matrix λI − A is obtained from the matrix A − λI by multiplying each
of its n rows by −1, and hence det(λI − A) = (−1)n det(A − λI ). In practice, it is most convenient
to work with A − λI in finding eigenvectors—this minimizes arithmetic—and when we come to
find chains of generalized eigenvectors in Section 1.2, it is (almost) essential to use A − λI , as using
λI − A would introduce lots of spurious minus signs.
5 −7
Example 1.5. Returning to the matrix A = of Example 1.2, we compute that det(λI −
2 −4
A) = λ2 − λ − 6 = (λ − 3)(λ + 2), so A has eigenvalues 3 and −2. Computation then shows that
7
the eigenspace E3 = Ker(A − 3I ) has basis , and that the eigenspace E−2 = Ker(A −
2
1
(−2)I ) has basis .
1
Definition 1.6. Let a be an eigenvalue of a matrix A. The algebraic multiplicity of the eigenvalue
a is alg-mult(a) = the multiplicity of a as a root of the characteristic polynomial det(λI − A). The
geometric multiplicity of the eigenvalue a is geom-mult(a) = the dimension of the eigenspace Ea .
It is common practice to use the word multiplicity (without a qualifier) to mean algebraic
multiplicity.
We have the following relationship between these two multiplicities.
1 ≤ geom-mult(a) ≤ alg-mult(a) .
Proof. By the definition of an eigenvalue, there is at least one eigenvector v with eigenvalue a, and
so Ea contains the nonzero vector v, and hence dim(Ea ) ≥ 1.
For the proof that geom-mult(a) ≤ alg-mult(a), see Lemma 1.12 in Appendix A. 2
Corollary 1.8. Let a be an eigenvalue of A and suppose that a has algebraic multiplicity 1. Then a also
has geometric multiplicity 1.
1 ≤ geom-mult(a) ≤ alg-mult(a) = 1 ,
so geom-mult(a) = 1. 2
1.1. THE DIAGONALIZABLE CASE 3
Remark 1.9. Let A be an n-by-n matrix.Then its characteristic polynomial det(λI − A) has degree
n. Since we are considering A to be defined over the complex numbers, we may apply the Fundamental
Theorem of Algebra, which states that an nth degree polynomial has n roots, counting multiplicities.
Hence, we see that, for any n-by-n matrix A, the sum of the algebraic multiplicities of the eigenvalues
of A is equal to n.
Proof. Let A have eigenvalues a1 , a2 , . . . , am . For each i between 1 and m, let si = geom-mult(ai )
and ti = alg-mult(ai ). Then, by Lemma 1.7, si ≤ ti for each i, and by Remark 1.9, m i=1 ti = n.
m m
Thus, if si = ti for each i, then i=1 si = n, while if si < ti for some i, then i=1 si < n. 2
Proposition 1.11. (1) Let a1 , a2 , . . . , am be distinct eigenvalues of A (i.e., ai = aj for i = j ). For each
i between 1 and m, let vi be an associated eigenvector. Then {v1 , v2 , . . . , vm } is a linearly independent set
of vectors.
(2) More generally, let a1 , a2 , . . . , am be distinct eigenvalues of A. For each i between 1 and m, let Si be
a linearly independent set of eigenvectors associated to ai . Then S = S1 ∪ . . . Sm is a linearly independent
set of vectors.
Proceed in this way, until at the last step we multiply by (A − a2 I ). We then obtain:
where v1 = c1,1 v1,1 + . . . + c1,i1 v1,i1 and v2 = c2,1 v2,1 + . . . + c2,i2 v2,i2 . But v1 is a vector in Ea1 ,
so Av1 = a1 v1 ; similarly, v2 is a vector in Ea2 , so Av2 = a2 v2 . Then, as in the proof of part (1),
Definition 1.12. Two square matrices A and B are similar if there is an invertible matrix P with
A = P BP −1 .
1.1. THE DIAGONALIZABLE CASE 5
Definition 1.13. A square matrix A is diagonalizable if A is similar to a diagonal matrix.
Theorem 1.14. Let A be an n-by-n matrix over the complex numbers. Then A is diagonalizable if
and only if, for each eigenvalue a of A, geom-mult(a) = alg-mult(a). In that case, A = P J P −1 where
J is a diagonal matrix whose entries are the eigenvalues of A, each appearing according to its algebraic
multiplicity, and P is a matrix whose columns are eigenvectors forming bases for the associated eigenspaces.
Proof. We give a proof by direct computation here. For a more conceptual proof, see Theorem 1.10
in Appendix A.
Let A have eigenvalues a1 , a2 , …, an . Here we do not insist that the ai ’s are distinct; rather,
each eigenvalue appears the same number of times as its algebraic multiplicity.Then J is the diagonal
matrix
J = j1 j2 . . . jn
P ji = ai vi
as well.
Thus, every column of AP is equal to the corresponding column of P J , so
AP = P J .
By Proposition 1.11, the columns of the square matrix P are linearly independent, so P is
invertible. Multiplying on the right by P −1 , we see that
A = P J P −1 ,
Corollary 1.15. Let A be an n-by-n matrix over the complex numbers all of whose eigenvalues are
distinct (i.e., whose characteristic polynomial has no repeated roots). Then A is diagonalizable.
1.2. THE GENERAL CASE 7
Proof. By hypothesis, for each eigenvalue a of A, alg-mult(a) = 1. But then, by Corollary 1.8, for
each eigenvalue a of A, geom-mult(a) = alg-mult(a), so the hypothesis of Theorem 1.14 is satisfied.
2
5 −7
Example 1.16. Let A be the matrix A = of Examples 1.2 and 1.5. Then, referring to
2 −4
Example 1.5, we see
−1
5 −7 7 1 3 0 7 1
= .
2 −4 2 1 0 −2 2 1
As we have indicated, we have developed this theory over the complex numbers, as JFC works
best over them. But there is an analog of our results over the real numbers—we just have to require
that all the eigenvalues of A are real. Here is the basic result on diagonalizability.
Theorem 1.17. Let A be an n-by-n real matrix. Then A is diagonalizable if and only if all the
eigenvalues of A are real numbers, and, for each eigenvalue a of A, geom-mult(a) = alg-mult(a). In that
case, A = P J P −1 where J is a diagonal matrix whose entries are the eigenvalues of A, each appearing
according to its algebraic multiplicity (and hence is a real matrix), and P is a real matrix whose columns
are eigenvectors forming bases for the associated eigenspaces.
Definition 2.1. A k-by-k Jordan block associated to the eigenvalue λ is a k-by-k matrix of the form
⎡ ⎤
λ 1
⎢ λ 1 ⎥
⎢ ⎥
⎢ λ 1 ⎥
⎢ ⎥
J =⎢ .. .. ⎥ .
⎢ . . ⎥
⎢ ⎥
⎣ λ 1⎦
λ
8 CHAPTER 1. JORDAN CANONICAL FORM
In other words, a Jordan block is a matrix with all the diagonal entries equal to each other, all
the entries immediately above the diagonal equal to 1, and all the other entries equal to 0.
Definition 2.2. A matrix J in Jordan Canonical Form ( JCF) is a block diagonal matrix
⎡ ⎤
J1
⎢ J2 ⎥
⎢ ⎥
⎢ ⎥
J =⎢ J3 ⎥
⎢ .. ⎥
⎣ . ⎦
J
Remark 2.3. Note that every diagonal matrix is a matrix in JCF, with each Jordan block a 1-by-1
block.
In order to understand and be able to use JCF, we must introduce a new concept, that of a
generalized eigenvector.
(A − λI )k (v) = 0
for some positive integer k, then v is a generalized eigenvector of A associated to the eigenvalue λ.
The smallest k with (A − λI )k (v) = 0 is the index of the generalized eigenvector v.
(A − λI )(v) = 0
(A)v = (λI )v
Av = λv
Eλ = {v | Av = λv} = {v | (A − λI )v = 0} .
Proposition 2.5. Let λ be an eigenvalue of the n-by-n matrix A of algebraic multiplicity m. Then, E˜λ
is a subspace of Cn of dimension m.
0 1 1
Example 2.6. Let A be the matrix A = . Then, as you can check, if u = , then
−4 4 2
(A − 2I )u = 0, so u is an eigenvector of A with associated eigenvalue 2 (and hence a generalized
1
eigenvector of index 1 of A with associated eigenvalue 2). On the other hand, if v = , then
0
(A − 2I )2 v = 0 but (A − 2I )v = 0, so v is a generalized eigenvector of index 2 of A with associated
eigenvalue 2.
In this case, as you can check, the vector u is a basis for the eigenspace E2 , so E2 = { cu | c ∈ C}
is one dimensional.
On the other hand, u and v are both generalized eigenvectors associated to the eigenvalue
2, and are linearly independent (the equation c1 u + c2 v = 0 only has the solution c1 = c2 = 0, as
you can readily check), so Ẽ2 has dimension at least 2. Since Ẽ2 is a subspace of C2 , it must have
dimension exactly 2, and Ẽ2 = C2 (and {u, v} is indeed a basis for C2 ).
vk−1 = (A − λI )vk .
We claim that vk−1 is a generalized eigenvector of index k − 1 associated to the eigenvalue λ.
To see this, note that
(A − λI )k−1 vk−1 = (A − λI )k−1 (A − λI )vk = (A − λI )k vk = 0
but
(A − λI )k−2 vk−1 = (A − λI )k−2 (A − λI )vk = (A − λI )k−1 vk = 0 .
Proceeding in this way, we may set
vk−2 = (A − λI )vk−1 = (A − λI )2 vk
vk−3 = (A − λI )vk−2 = (A − λI )2 vk−1 = (A − λI )3 vk
..
.
v1 = (A − λI )v2 = · · · = (A − λI )k−1 vk
10 CHAPTER 1. JORDAN CANONICAL FORM
and note that each vi is a generalized eigenvector of index i associated to the eigenvalue λ. A
collection of generalized eigenvectors obtained in this way gets a special name:
then {v1 , . . . , vk } is called a chain of generalized eigenvectors of length k. The vector vk is called
the top of the chain and the vector v1 (which is an ordinary eigenvector) is called the bottom of the
chain.
1
Example 2.8. Let us return to Example 2.6. We saw there that v = is a generalized eigenvector
0
0 1 1
of index 2 of A = associated to the eigenvalue 2. Let us set v2 = v = . Then v1 =
−4 4 0
−2
(A − 2I )v2 = is a generalized eigenvector of index 1 (i.e., an ordinary eigenvector), and
−4
{v1 , v2 } is a chain of length 2.
Remark 2.9. It is important to note that a chain of generalized eigenvectors {v1 , . . . , vk } is entirely
determined by the vector vk at the top of the chain. For once we have chosen vk , there are no other
choices to be made: the vector vk−1 is determined by the equation vk−1 = (A − λI )vk ; then the
vector vk−2 is determined by the equation vk−2 = (A − λI )vk−1 , etc.
With this concept in hand, let us return to JCF. As we have seen, a matrix J in JCF has a
number of blocks J1 , J2 , . . . , J , called Jordan blocks, along the diagonal. Let us begin our analysis
with the case when J consists of a single Jordan block. So suppose J is a k-by-k matrix
⎡ ⎤
λ 1
⎢ λ 1 0 ⎥
⎢ ⎥
⎢ λ 1 ⎥
⎢ ⎥
J =⎢ .. .. ⎥ .
⎢ . . ⎥
⎢ ⎥
⎣ 0 λ 1⎦
λ
1.2. THE GENERAL CASE 11
Then,
⎡ ⎤
0 1
⎢ 0 1 ⎥
⎢ ⎥
⎢ 0 1 ⎥
⎢ ⎥
J − λI = ⎢ .. .. ⎥ .
⎢ . . ⎥
⎢ ⎥
⎣ 0 1⎦
0
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0 0 0
⎢0⎥ ⎢1⎥ ⎢0⎥ ⎢0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
Let e1 = ⎢0⎥, e2 = ⎢0⎥, e3 = ⎢1⎥, …, ek = ⎢0⎥ .
⎢.⎥ ⎢.⎥ ⎢.⎥ ⎢.⎥
⎣ .. ⎦ ⎣ .. ⎦ ⎣ .. ⎦ ⎣ .. ⎦
0 0 0 1
Then direct calculation shows:
(J − λI )ek = ek−1
(J − λI )ek−1 = ek−2
..
.
(J − λI )e2 = e1
(J − λI )e1 = 0
and so we see that {e1 , . . . , ek } is a chain of generalized eigenvectors. We also note that {e1 , . . . , ek }
is a basis for Ck , and so
E˜λ = Ck .
We first see that the situation is very analogous when we consider any k-by-k matrix with a
single chain of generalized eigenvectors of length k.
Proposition 2.10. Let {v1 , . . . , vk } be a chain of generalized eigenvectors of length k associated to the
eigenvalue λ of a matrix A. Then {v1 , . . . , vk } is linearly independent.
c1 v1 + c2 v2 + · · · + ck−1 vk−1 + ck vk = 0 .
Now let us multiply this equation on the left by (A − λI )k−1 . Then we obtain the equation
ck v1 = 0 .
c1 v1 + c2 v2 + · · · + ck−1 vk−1 = 0 .
Repeat the same argument, this time multiplying by (A − λI )k−2 instead of (A − λI )k−1 .
Then we obtain the equation
ck−1 v1 = 0 ,
and, since v1 = 0, this shows that ck−1 = 0 as well. Keep going to get
c1 = c2 = · · · = ck−1 = ck = 0 ,
Theorem 2.11. Let A be a k-by-k matrix and suppose that Ck has a basis {v1 , . . . , vk } consisting of a
single chain of generalized eigenvectors of length k associated to an eigenvalue a. Then
A = P J P −1
where ⎡ ⎤
a 1
⎢ a 1 ⎥
⎢ ⎥
⎢ a 1 ⎥
⎢ ⎥
J =⎢ .. .. ⎥
⎢ . . ⎥
⎢ ⎥
⎣ a 1⎦
a
is a matrix consisting of a single Jordan block and
P = v1 v2 . . . vk
Let P be the given matrix. We will first show by direct computation that AP = P J .
It will be convenient to write
J = j1 j2 . . . jk
Avi = (A − aI + aI )vi
= (A − aI )vi + aI vi
= vi−1 + avi for i > 1, = avi for i = 1 .
as well.
14 CHAPTER 1. JORDAN CANONICAL FORM
Thus, every column of AP is equal to the corresponding column of P J , so
AP = P J .
But Proposition 2.10 shows that the columns of P are linearly independent, so P is invertible.
Multiplying on the right by P −1 , we see that
A = P J P −1 .
0 1
Example 2.12. Applying Theorem 2.11 to the matrix A = of Examples 2.6 and 2.8, we
−4 4
see that
−1
0 1 −2 1 2 1 −2 1
= .
−4 4 −4 0 0 2 −4 0
Here is the key theorem to which we have been heading. This theorem is one of the most
important (and useful) theorems in linear algebra.
Theorem 2.13. Let A be any square matrix defined over the complex numbers. Then A is similar to a
matrix in Jordan Canonical Form. More precisely, A = P J P −1 , for some matrix J in Jordan Canonical
Form. The diagonal entries of J consist of eigenvalues of A, and P is an invertible matrix whose columns
are chains of generalized eigenvectors of A.
Proof. (Rough outline) In general, the JCF of a matrix A does not consist of a single block, but will
have a number of blocks, of varying sizes and associated to varying eigenvalues.
But in this situation we merely have to “assemble” the various blocks (to get the matrix J )
and the various chains of generalized eigenvectors (to get a basis and hence the matrix P ). Actually,
the word “merely” is a bit misleading, as the argument that we can do so is, in fact, a subtle one, and
we shall not give it here. 2
In lieu of proving Theorem 2.13, we shall give a number of examples that illustrate the
situation. In fact, in order to avoid complicated notation we shall merely illustrate the situation for
2-by-2 and 3-by-3 matrices.
Theorem 2.14. Let A be a 2-by-2 matrix. Then one of the following situations applies:
1.2. THE GENERAL CASE 15
(i) A has two eigenvalues, a and b, each of algebraic multiplicity 1. Let u be an eigenvector associated to
the eigenvalue a and let v be an eigenvector associated to the eigenvalue b. Then A = P J P −1 with
a 0
J = and P = u v .
0 b
(Note, in this case, A is diagonalizable.)
(ii) A has a single eigenvalue a of algebraic multiplicity 2.
(a) A has two linearly independent eigenvectors u and v.
Then A = P J P −1 with
a 0
J = and P = u v .
0 a
(Note, in this case, A is diagonalizable. In fact, in this case Ea = C2 and A itself is the matrix
a 0
.)
0 a
(b) A has a single chain {v1 , v2 } of generalized eigenvectors. Then A = P J P −1 with
a 1
J = and P = v1 v2 .
0 a
Theorem 2.15. Let A be a 3-by-3 matrix. Then one of the following situations applies:
(i) A has three eigenvalues, a, b, and c, each of algebraic multiplicity 1. Let u be an eigenvector
associated to the eigenvalue a, v be an eigenvector associated to the eigenvalue b, and w be an
eigenvector associated to the eigenvalue c. Then A = P J P −1 with
⎡ ⎤
a 0 0
⎣
J = 0 b 0 ⎦ and P = u v w .
0 0 c
(Note, in this case, A is diagonalizable.)
(ii) A has an eigenvalue a of algebraic multiplicity 2 and an eigenvalue b of algebraic multiplicity 1.
(a) A has two independent eigenvectors, u and v, associated to the eigenvalue a. Let w be an eigenvector
associated to the eigenvalue b. Then A = P J P −1 with
⎡ ⎤
a 0 0
⎣
J = 0 a 0 ⎦ and P = u v w .
0 0 b
(Note, in this case, A is diagonalizable.)
16 CHAPTER 1. JORDAN CANONICAL FORM
(b) A has a single chain {u1 , u2 } of generalized eigenvectors associated to the eigenvalue a. Let v be an
eigenvector associated to the eigenvalue b. Then A = P J P −1 with
⎡ ⎤
a 1 0
⎣
J = 0 a 0 ⎦ and P = u1 u2 v .
0 0 b
(Note, in this case, A is diagonalizable. In fact, in this case Ea = C3 and A itself is the matrix
⎡ ⎤
a 0 0
⎣0 a 0⎦.)
0 0 a
(b) A has a chain {u1 , u2 } of generalized eigenvectors and an eigenvector v with {u1 , u2 , v} linearly
independent. Then A = P J P −1 with
⎡ ⎤
a 1 0
J = ⎣0 a 0⎦ and P = u1 u2 v .
0 0 a
Remark 2.16. Suppose that A has JCF J = aI , a scalar multiple of the identity matrix. Then
A = P J P −1 = P (aI )P −1 = a(P I P −1 ) = aI = J .This justifies the parenthetical remark in The-
orems 2.14 (ii) (a) and 2.15 (iii) (a).
Remark 2.17. Note that Theorems 2.14 (i), 2.14 (ii) (a), 2.15 (i), 2.15 (ii) (a), and 2.15 (iii) (a) are
all special cases of Theorem 1.14, and in fact Theorems 2.14 (i) and 2.15 (i) are both special cases
of Corollary 1.15.
1.2. THE GENERAL CASE 17
Now we would like to apply Theorems 2.14 and 2.15. In order to do so, we need to have an
effective method to determine which of the cases we are in, and we give that here (without proof ).
Eλi = {v | (A − λI )i (v) = 0}
= Ker((A − λI )i ) .
Note that Eλi consists of generalized eigenvectors of index at most i (and the 0 vector), and is
a subspace. Note also that
Eλ = Eλ1 ⊆ Eλ2 ⊆ . . . ⊆ E˜λ .
In general, the JCF of A is determined by the dimensions of all the spaces Eλi , but this
determination can be a bit complicated. For eigenvalues of multiplicity at most 3, however, the
situation is simpler—we need only consider the eigenspaces Eλ .This is a consequence of the following
general result.
Proposition 2.19. Let λ be an eigenvalue of A. Then the number of blocks in the JCF of A corresponding
to λ is equal to dim Eλ , i.e., to the geometric multiplicity of λ.
Proof. (Outline) Suppose there are such blocks. Since each block corresponds to a chain of gener-
alized eigenvectors, there are such chains. Now the bottom of the chain is an (ordinary) eigenvector,
so we get eigenvectors in this way. It can be shown that these eigenvectors are always linearly
independent and that they always span Eλ , i.e., that they are a basis of Eλ . Thus, Eλ has a basis
consisting of vectors, so dim Eλ = . 2
We can now determine the JCF of 1-by-1, 2-by-2, and 3-by-3 matrices, using the following
consequences of this proposition.
Corollary 2.20. Let λ be an eigenvalue of A of algebraic multiplicity 1. Then dim Eλ1 = 1, i.e., a has
geometric multiplicity 1, and the submatrix of the JCF of A corresponding to the eigenvalue λ is a single
1-by-1 block.
Corollary 2.21. Let λ be an eigenvalue of A of algebraic multiplicity 2. Then there are the following
possibilities:
(a) dim Eλ1 = 2, i.e., a has geometric multiplicity 2. In this case, the submatrix of the JCF of A
corresponding to the eigenvalue λ consists of two 1-by-1 blocks.
18 CHAPTER 1. JORDAN CANONICAL FORM
(b) dim Eλ1 = 1, i.e., a has geometric multiplicity 1. Also, dim Eλ2 = 2. In this case, the submatrix of
the JCF of A corresponding to the eigenvalue λ consists of a single 2-by-2 block.
Corollary 2.22. Let λ be an eigenvalue of A of algebraic multiplicity 3. Then there are the following
possibilities:
(a) dim Eλ1 = 3, i.e., a has geometric multiplicity 3. In this case, the submatrix of the JCF of A corre-
sponding to the eigenvalue λ consists of three 1-by-1 blocks.
(b) dim Eλ1 = 2, i.e., a has geometric multiplicity 2. Also, dim Eλ2 = 3. In this case, the submatrix of
the Jordan Canonical Form of A corresponding to the eigenvalue λ consists of a 2-by-2 block and a
1-by-1 block.
(c) dim Eλ1 = 1, i.e., a has geometric multiplicity 1. Also, dim Eλ2 = 2, and dim Eλ3 = 3. In this case,
the submatrix of the Jordan Canonical Form of A corresponding to the eigenvalue λ consists of a
single 3-by-3 block.
⎡ ⎤
3 1 1
Example 2.24. A = ⎣2 4 2⎦ .
1 1 3
1.2. THE GENERAL CASE 19
⎡ ⎤
2 1 1
Example 2.25. A = ⎣ 2 1 −2⎦ .
−1 0 −2
A has characteristic polynomial det (λI − A) = (λ + 1)2 (λ − 3). Thus, A has an eigenvalue
−1 of multiplicity 2 and an eigenvalue
⎧⎡ 3 ⎤ of⎫multiplicity 1. Computation shows that the eigenspace
⎨ −1 ⎬
E−1 = Ker(A − (−I )) has basis ⎣ 2 ⎦ so dim E−1 = 1 and we are in the situation of Corol-
⎩ ⎭
1
⎧⎡ ⎤ ⎡ ⎤⎫
⎨ −1 0 ⎬
lary 2.21 (b). Then we further compute that E−1 2 = Ker((A − (−I ))2 ) has basis ⎣ 2 ⎦ , ⎣0⎦ ,
⎩ ⎭
0 1
therefore is two-dimensional, as we expect. More to the point, we may choose any generalized eigen-
2 that is not in E 1 , as the top of a chain. We choose u =
vector of index 2, i.e., any vector in E−1
⎡ ⎤ −1
2
⎡ ⎤
0 1
⎣0⎦ , and then we have u1 = (A − (−I ))u2 = ⎣−2⎦ , and {u1 , u2 } form a chain.
1 −1
⎧ ⎡ ⎤⎫
⎨ −5 ⎬
We also compute that, for the eigenvalue 3, the eigenspace E3 has basis v = ⎣−6⎦ .
⎩ ⎭
1
Hence, we see that
⎡ ⎤ ⎡ ⎤⎡ ⎤⎡ ⎤−1
2 1 1 1 0 −5 −1 1 0 1 0 −5
⎣ 2 1 −2⎦ = ⎣−2 0 −6⎦ ⎣ 0 −1 0⎦ ⎣−2 0 −6⎦ .
−1 0 2 −1 1 1 0 0 3 −1 1 1
20 CHAPTER 1. JORDAN CANONICAL FORM
⎡ ⎤
2 1 1
Example 2.26. A = ⎣−2 −1 −2⎦ .
1 1 2
A has characteristic polynomial det (λI − A) = (λ − 1)3 , so A has ⎧⎡one ⎤eigenvalue 1 of
⎡ ⎤⎫
⎨ −1 −1 ⎬
multiplicity three. Computation shows that E1 = Ker(A − I ) has basis ⎣ 0 ⎦ , ⎣ 1 ⎦ , so
⎩ ⎭
1 0
dim E1 = 2 and we are in the situation of Corollary 2.22⎧ (b). Computation⎫then shows that
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎨ 1 0 0 ⎬
dim E12 = 3 (i.e., (A − I )2 = 0 and E12 is all of C3 ) with basis ⎣0⎦ , ⎣1⎦ , ⎣0⎦ .We may choose
⎩ ⎭
0 0 1
⎡ ⎤
1
u2 to be any vector in E12 that is not in E11 , and we shall choose u2 = ⎣0⎦ . Then u1 = (A − I )u2 =
0
⎡ ⎤
1
⎣−2⎦ , and {u1 , u2 } form a chain. For the third vector, v, we may choose any vector in E1 such that
1
⎡ ⎤
−1
{u1 , v} is linearly independent. We choose v = ⎣ 0 ⎦ . Hence, we see that
1
⎡ ⎤ ⎡ ⎤⎡ ⎤⎡ ⎤−1
2 1 1 1 1 −1 1 1 0 1 1 −1
⎣−2 −1 2⎦ = ⎣−2 0 0 ⎦ ⎣0 1 0⎦ ⎣−2 0 0 ⎦ .
1 1 2 1 0 1 0 0 1 1 0 1
⎡ ⎤
5 0 1
Example 2.27. A = ⎣ 1 1 0⎦ .
−7 1 0
A has characteristic polynomial det (λI − A) = (λ − 2)3 , so A
⎧has one eigenvalue 2 of multi-
⎡ ⎤⎫
⎨ −1 ⎬
plicity three. Computation shows that E2 = Ker(A − 2I ) has basis ⎣−1⎦ , so dim E21 = 1 and
⎩ ⎭
3
we are in the situation of Corollary 2.22 (c). Then computation shows that E22 = Ker((A − 2I )2 )
⎧⎡ ⎤ ⎡ ⎤⎫ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎨ −1 −1 ⎬ −1 −1 −1
has basis ⎣ ⎦
0 , 2⎣ ⎦ . (Note that −1 = 3/2 0 + 1/2 2 ⎦.) Computation then
⎣ ⎦ ⎣ ⎦ ⎣
⎩ ⎭
2 0 3 2 0
1.2. THE GENERAL CASE 21
⎧⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎫
⎨ 1 0 0 ⎬
shows that dim E2 = 3 (i.e., (A − 2I ) = 0 and E2 is all of C ) with basis
3 3 3 3 ⎣ 0 , 1 , 0⎦ .
⎦ ⎣ ⎦ ⎣
⎩ ⎭
0 0 1
⎡ ⎤
1
We may choose u3 to be any vector in C3 that is not in E22 , and we shall choose u3 = ⎣0⎦ . Then
0
⎡ ⎤ ⎡ ⎤
3 2
u2 = (A − 2I )u3 = ⎣ 1 ⎦ and u1 = (A − 2I )u2 = ⎣ 2 ⎦ , and then {u1 , u2 , u3 } form a chain.
−7 −6
Hence, we see that
⎡ ⎤ ⎡ ⎤⎡ ⎤⎡ ⎤−1
5 0 1 2 3 1 2 1 0 2 3 1
⎣ 1 1 0⎦ = ⎣ 2 1 0⎦ ⎣0 2 1⎦ ⎣ 2 1 0⎦ .
−7 1 0 −6 −7 0 0 0 2 −6 −7 0
As we have mentioned, we need to work over the complex numbers in order for the theory
of JCF to fully apply. But there is an analog over the real numbers, and we conclude this section by
stating it.
Theorem 2.28. Let A be a real square matrix (i.e., a square matrix with all entries real numbers), and
suppose that all of the eigenvalues of A are real numbers. Then A is similar to a real matrix in Jordan
Canonical Form. More precisely, A = P J P −1 with P and J real matrices, for some matrix J in Jordan
Canonical Form. The diagonal entries of J consist of eigenvalues of A, and P is an invertible matrix whose
columns are chains of generalized eigenvectors of A.
75 56
1. A = , det(λI − A) = (λ − 3)(λ − 5).
−90 −67
−50 99
2. A = , det(λI − A) = (λ + 6)(λ + 5).
−20 39
−18 9
3. A = , det(λI − A) = (λ − 3)2 .
−49 24
22 CHAPTER 1. JORDAN CANONICAL FORM
1 1
4. A = , det(λI − A) = (λ − 5)2 .
−16 9
2 1
5. A = , det(λI − A) = (λ − 7)2 .
−25 12
−15 9
6. A = , det(λI − A) = λ2 .
−25 15
⎡ ⎤
1 0 0
7. A = ⎣1 2 −3⎦, det(λI − A) = (λ + 1)(λ − 1)(λ − 3).
1 −1 0
⎡ ⎤
3 0 2
8. A = ⎣1 3 1⎦, det(λI − A) = (λ − 1)(λ − 2)(λ − 4).
0 1 1
⎡ ⎤
5 8 16
9. A = ⎣ 4 1 8 ⎦, det(λI − A) = (λ + 3)2 (λ − 1).
−4 −4 −11
⎡ ⎤
4 2 3
10. A = ⎣−1 1 −3⎦, det(λI − A) = (λ − 3)2 (λ − 8).
2 4 9
⎡ ⎤
5 2 1
11. A = ⎣−1 2 −1⎦, det(λI − A) = (λ − 4)2 (λ − 2).
−1 −2 3
⎡ ⎤
8 −3 −3
12. A = ⎣ 4 0 −2⎦, det(λI − A) = (λ − 2)2 (λ − 7).
−2 1 3
⎡ ⎤
−3 1 −1
13. A = ⎣−7 5 −1⎦, det(λI − A) = (λ + 2)2 (λ − 4).
−6 6 −2
⎡ ⎤
3 0 0
14. A = ⎣ 9 −5 −18⎦, det(λI − A) = (λ − 3)2 (λ − 4).
−4 4 12
1.2. THE GENERAL CASE 23
⎡ ⎤
−6 9 0
15. A = ⎣−6 6 −2⎦, det(λI − A) = λ2 (λ − 3).
9 −9 3
⎡ ⎤
−18 42 168
16. A = ⎣ 1 −7 −40⎦, det(λI − A) = (λ − 3)2 (λ + 4).
−2 6 27
⎡ ⎤
−1 1 −1
17. A = ⎣−10 6 −5⎦, det(λI − A) = (λ − 1)3 .
−6 3 −2
⎡ ⎤
0 −4 1
18. A = ⎣2 −6 1⎦, det(λI − A) = (λ + 2)3 .
4 −8 0
⎡ ⎤
−4 1 2
19. A = ⎣−5 1 3⎦, det(λI − A) = λ3 .
−7 2 3
⎡ ⎤
−4 −2 5
20. A = ⎣−1 −1 1⎦, det(λI − A) = (λ + 1)3 .
−2 −1 2
24
25
CHAPTER 2
Y = (P J P −1 )Y
Y = P J (P −1 Y )
P −1 Y = J (P −1 Y )
(P −1 Y ) = J (P −1 Y ) .
Z = J Z
Y = PZ
is the solution to our original system.
Examining this strategy, we see that we already know how to carry out Step 1, and also that
Step 3 is very easy—it is just matrix multiplication. Thus, the key to success here is being able to
carry out Step 2. This is where JCF comes in. As we shall see, it is (relatively) easy to solve Z = J Z
when J is a matrix in JCF.
26 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
You will note that throughout this section, in solving Z = J Z, we write the solution as
Z = MZ C, where MZ is a matrix of functions, called the fundamental matrix of the system, and
C is a vector of arbitrary constants. The reason for this will become clear later. (See Remarks 1.12
and 1.14.)
Although it is not logically necessary—we may regard a diagonal matrix as a matrix in JCF
in which all the Jordan blocks are 1-by-1 blocks—it is illuminating to handle the case when J is
diagonal first. Here the solution is very easy.
Proof. Multiplying out, we see that the system Z = J Z is just the system
⎡ ⎤ ⎡ ⎤
z1 a1 z1
⎢ ⎥ ⎢ ⎥
⎢z2 ⎥ ⎢a2 z2 ⎥
⎢ ⎥ ⎢ ⎥
⎢.⎥=⎢ . ⎥ .
⎢ .. ⎥ ⎢ .. ⎥
⎣ ⎦ ⎣ ⎦
zk a k zk
2.1. HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 27
But this system is “uncoupled”, i.e., the equation for zi
only involves zi and none of the
other functions. Now this equation is very familiar. In general, the differential equation z = az has
solution z = ceax , and applying that here we find that Z = J Z has solution
⎡ a1 x ⎤
c1 e
⎢c2 ea2 x ⎥
⎢ ⎥
Z=⎢ . ⎥ ,
⎣ .. ⎦
ck eak x
and so Y = P Z = P MZ C, i.e.,
7 1 e3x c10
Y =
2 1 e−2x c2
0
3x
7e e−2x c1
=
2e3x e−2x c2
7c1 e3x + c2 e−2x
= .
2c1 e3x + c2 e−2x
and so Y = P Z = P MZ C, i.e.,
⎡ ⎤ ⎡ −x ⎤⎡ ⎤
1 0 −1 e 0 0 c1
⎣
Y = 0 −1 −1 ⎦ ⎣ 0 1 0 ⎦ ⎣c2 ⎦
1 1 1 0 0 e 2x c3
⎡ −x ⎤ ⎡ ⎤
e 0 −e2x c1
= ⎣ 0 −1 −e2x ⎦ ⎣c2 ⎦
e−x 1 e2x c3
⎡ −x ⎤
c1 e − c3 e2x
=⎣ −c2 − c3 e2x ⎦ .
c1 e−x + c2 + c3 e2x
We now see how to use JCF to solve systems Y = AY where the coefficient matrix A is not
diagonalizable.
The key to understanding systems is to investigate a system Z = J Z where J is a matrix
consisting of a single Jordan block. Here the solution is not as easy as in Theorem 1.1, but it is still
not too hard.
Proof. We will prove this in the cases k = 1, 2, and 3, which illustrate the pattern. As you will see,
the proof is a simple application of the standard technique for solving first-order linear differential
equations.
The case k = 1: Here we are considering the system
[z1 ] = [a][z1 ]
z1 = az1 .
z1 = c1 eax ,
z1 = az1 + z2
z2 = az2 .
30 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
We recognize the second equation as having the solution
z2 = c2 eax
and recognize that this differential equation has integrating factor e−ax . Multiplying by this factor,
we find
so
z1 = eax (c1 + c2 x) .
Thus, our solution is
z1 = eax (c1 + c2 x)
z2 = eax c2 ,
z1 = az1 + z2
z2 = az2 + z3
z3 = az3 .
2.1. HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 31
If we just concentrate on the last two equations, we see we are in the k = 2 case. Referring to
that case, we see that our solution is
z2 = eax (c2 + c3 x)
z3 = eax c3 .
and recognize that this differential equation has integrating factor e−ax . Multiplying by this factor,
we find
so
z1 = eax (c1 + c2 x + c3 (x 2 /2)) .
Thus, our solution is
Remark 1.5. Suppose that Z = J Z where J is a matrix in JCF but one consisting of several blocks,
not just one block. We can see that this systems decomposes into several systems, one corresponding
to each block, and that these systems are uncoupled, so we may solve them each separately, using
Theorem 1.4, and then simply assemble these individual solutions together to obtain a solution of
the general system.
32 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
We now illustrate this (confining our illustrations to the case that A is not diagonalizable, as
we have already illustrated the diagonalizable case).
and so Y = P Z = P MZ C, i.e.,
−2 1 2x 1 x c1
Y = e
−4 0 0 1 c2
2x
−2 1 e xe2x c1
=
−4 0 0 e2x c2
−2e2x −2xe2x + e2x c1
=
−4e2x −4xe2x c2
(−2c1 + c2 )e2x − 2c2 xe2x
= .
−4c1 e2x − 4c2 xe2x
and so Y = P Z = P MZ C, i.e.,
⎡ ⎤ ⎡ −x ⎤⎡ ⎤
1 0 −5 e xe−x 0 c1
Y = ⎣−2 0 −6⎦ ⎣ 0 e−x 0 ⎦ ⎣c2 ⎦
−1 1 1 0 0 e3x c3
⎡ −x ⎤ ⎡ ⎤
e xe−x −5e3x c1
= ⎣−2e−x −2xe−x −6e3x ⎦ ⎣c2 ⎦
−e −x −x
−xe + e −x e3x c3
⎡ ⎤
c1 e−x + c2 xe−x − 5c3 e3x
= ⎣ −2c1 e−x − 2c2 xe−x − 6c3 e3x ⎦ .
(−c1 + c2 )e−x − c2 xe−x + c3 e3x
and so Y = P Z = P MZ C, i.e.,
⎡ ⎤⎡ x ⎤⎡ ⎤
1 1 1 e xex 0 c1
⎣
Y = −2 0 0 ⎦ ⎣ 0 ex 0 ⎦ ⎣c2 ⎦
1 0 1 0 0 e x c3
34 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
⎡ ⎤⎡ ⎤
ex xex + ex ex c1
⎣
= −2e x −2xe x 0 ⎦ ⎣c2 ⎦
e x xe x e x c3
⎡ ⎤
(c1 + c2 + c3 )ex + c2 xex
= ⎣−2c1 ex − 2c2 xex ⎦ .
(c1 + c3 )e x + c2 xex
and so Y = P Z = P MZ C, i.e.,
⎡ ⎤ ⎡ 2x ⎤⎡ ⎤
2 3 1 e xe2x (x 2 /2)e2x c1
Y =⎣ 2 1 0⎦ ⎣ 0 e2x xe2x ⎦ ⎣c2 ⎦
−6 −7 0 0 0 e2x c3
⎡ 2x ⎤⎡ ⎤
2e 2xe2x + 3e2x x 2 e2x + 3xe2x + e2x c1
⎣
= 2e 2x 2xe + e
2x 2x x e + xe
2 2x 2x ⎦ ⎣ c2 ⎦
−6e 2x −6xe − 7e
2x 2x −3x e − 7xe
2 2x 2x c3
⎡ ⎤
(2c1 + 3c2 + c3 )e2x + (2c2 + 3c3 )xe2x + c3 x 2 e2x
= ⎣ (2c1 + c2 )e2x + (2c2 + c3 )xe2x + c3 x 2 e2x ⎦ .
(−6c1 − 7c2 )e 2x + (−6c2 − 7c3 )xe2x − 3c3 x 2 e2x
2.1. HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 35
We conclude this section by showing how to solve initial value problems. This is just one more
step, given what we have already done.
In Example 1.6, we saw that this system has the general solution
(−2c1 + c2 )e2x − 2c2 xe2x
Y = .
−4c1 e2x − 4c2 xe2x
with solution
c1 2
= .
c2 7
Substituting these values in the above matrix gives
2x
3e − 14xe2x
Y = .
−8e2x − 28te2x
In Example 1.8, we saw that this system has the general solution
⎡ −x ⎤
c1 e + c2 xe−x − 5c3 xe3x
Y = ⎣ −2c1 e−x − 2c2 xe−x − 6c3 e3x ⎦ .
(−c1 + c2 )e−x − c2 xe−x + c3 e3x
Remark 1.12. There is a variant on our method of solving systems or initial value problems.
Z = J Z, Z(0) = Z0 ,
Z(x) = MZ (x)Z0 .
Now suppose we wish to solve the system Y = AY . Then, if A = P J P −1 , we have seen that
this system has solution Y = P Z = P MZ C. Let us manipulate this a bit:
Y = P MZ C = P MZ I C = P MZ (P −1 P )C
= (P MZ P −1 )(P C) .
Now let us set MY = P MZ P −1 , and also let us set = P C. Note that MY is still a matrix of
functions, and that is still a vector of arbitrary constants (since P is an invertible constant matrix
and C is a vector of arbitrary constants). Thus, with this notation, we see that
Y = AY has solution Y = MY .
2.1. HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 37
Now suppose we wish to solve the initial value problem
Y = AY, Y (0) = Y0 .
Rewriting the above solution of Y = AY to explicitly include the independent variable, we see that
we have
Y (x) = MY (x)
and, in particular,
Y0 = Y (0) = MY (0) = P MZ (0)P −1 = P I P −1 = ,
so we see that
Y = AY, Y (0) = Y0 has solution Y (x) = MY (x)Y0 .
This variant method has pros and cons. It is actually less effective than our original method
for solving a single initial value problem (as it requires us to compute P −1 and do some extra
matrix multiplication), but it has the advantage of expressing the solution directly in terms of the
initial conditions. This makes it more effective if the same system Y = AY is to be solved for a
variety of initial conditions. Also, as we see from Remark 1.14 below, it is of considerable theoretical
importance.
Remark 1.14. In Section 2.4 we will define the matrix exponential, and, with this definition,
MZ (x) = eJ x and MY (x) = P MZ (x)P −1 = eAx .
−50 99 7
2. A = and Y0 = .
−20 39 3
−18 9 41
3. A = and Y0 = .
−49 24 98
1 1 7
4. A = and Y0 = .
−16 9 16
2 1 −10
5. A = and Y0 = .
−25 12 −75
−15 9 50
6. A = and Y0 = .
−25 15 100
2.1. HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 39
⎡ ⎤ ⎡ ⎤
1 0 0 6
7. A = ⎣1 2 −3⎦ and Y0 = ⎣−10⎦.
1 −1 0 10
⎡ ⎤ ⎡ ⎤
3 0 2 0
⎦
8. A = 1 3 1 and Y0 = 3⎦.
⎣ ⎣
0 1 1 3
⎡ ⎤ ⎡ ⎤
5 8 16 0
9. A = ⎣ 4 1 8 ⎦ and Y0 = ⎣ 2 ⎦.
−4 −4 −11 −1
⎡ ⎤ ⎡ ⎤
4 2 3 3
10. A = ⎣−1 1 −3⎦ and Y0 = ⎣2⎦.
2 4 9 1
⎡ ⎤ ⎡ ⎤
5 2 1 −3
11. A = ⎣−1 2 −1⎦ and Y0 = ⎣ 2 ⎦.
−1 −2 3 9
⎡ ⎤ ⎡ ⎤
8 −3 −3 5
12. A = ⎣ 4 0 −2⎦ and Y0 = ⎣8⎦.
−2 1 3 7
⎡ ⎤ ⎡ ⎤
−3 1 −1 −1
⎦
13. A = −7 5 −1 and Y0 = 3 ⎦.
⎣ ⎣
−6 6 −2 6
⎡ ⎤ ⎡ ⎤
3 0 0 2
14. A = ⎣ 9 −5 −18⎦ and Y0 = ⎣−1⎦.
−4 4 12 1
40 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
⎡ ⎤ ⎡ ⎤
−6 9 0 1
15. A = ⎣−6 6 −2⎦ and Y0 = ⎣ 3 ⎦.
9 −9 3 −6
⎡ ⎤ ⎡ ⎤
−18 42 168 7
16. A = ⎣ 1 ⎦
−7 −40 and Y0 = −2⎦.
⎣
−2 6 27 1
⎡ ⎤ ⎡ ⎤
−1 1 −1 3
17. A = ⎣−10 6 −5⎦ and Y0 = ⎣10⎦.
−6 3 2 18
⎡ ⎤ ⎡ ⎤
0 −4 1 2
18. A = ⎣2 −6 1⎦ and Y0 = ⎣5⎦.
4 −8 0 8
⎡ ⎤ ⎡ ⎤
−4 1 2 6
19. A = ⎣−5 1 3⎦ and Y0 = ⎣11⎦.
−7 2 3 9
⎡ ⎤ ⎡ ⎤
−4 −2 5 9
20. A = ⎣−1 −1 1⎦ and Y0 = ⎣5⎦.
−2 −1 2 8
ez = 1 + z + z2 /2! + z3 /3! + . . . .
Lemma 2.3. Let A be a matrix with real entries, and let v be an eigenvector of A with associated
eigenvalue λ. Then v is an eigenvector of A with associated eigenvalue λ.
Proof. We have that Av = λv, by hypothesis. Let us take the complex conjugate of each side of this
equation. Then
Av = λv,
Av = λv,
Av = λv (as A = A since all the entries of A are real) ,
as claimed. 2
42 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
Now for our example.
We continue as before, but now we use F to denote a vector of arbitrary constants. (This is just for
neatness. Our constants will change, as you will see, and we will use the vector C to denote our final
constants, as usual.) Then Z = J Z has solution
(3+4i)x (3+4i)x
e 0 f1 f e
Z= (3−4i)x = MZ F = 1 (3−4i)x
0 e f2 f2 e
and so Y = P Z = P MZ F , i.e.,
−1 + 4i −1 − 4i e(3+4i)x 0 f1
Y =
1 1 0 e(3−4i)x f2
(3+4i)x −1 + 4i (3−4i)x −1 − 4i
= f1 e + f2 e .
1 1
Now we want our differential equation to have real solutions, and in order for this to be the
case, it turns out that we must have f2 = f1 . Thus, we may write our solution as
(3+4i)x −1 + 4i (3−4i)x −1 − 4i
Y = f1 e + f1 e
1 1
(3+4i)x −1 + 4i −1 + 4i
= f1 e + f1 e (3+4i)x ,
1 1
and so we obtain
e3x (− cos(4x) − 4 sin(4x)) e3x (4 cos(4x) − sin(4x)) c1
Y =
e3x cos(4x) e3x sin(4x) c2
(−c1 + 4c2 )e cos(4x) + (−4c1 − c2 )e sin(4x)
3x 3x
= .
c1 e3x cos(4x) + c2 e3x sin(4x)
Lemma 2.5. Consider the system Y = AY , where A is a matrix with real entries. Let this system have
general solution of the form
λ1 x
e 0 f1 f1
Y = P MZ F = v1 v1 1 x
= e v1 e 1 v1
λ λ x ,
0 e λ1 x f1 f1
where f1 is an arbitrary complex constant. Then this system also has general solution of the form
c1
Y = Re(eλ1 x v1 ) Im(eλ1 x v1 ) ,
c2
Proof. First note that for any complex number z = x + iy, x = Re(z) = 21 (z + z) and y = Im(z) =
2i (z − z), and similarly, for any complex vector.
1
44 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
Then
−1 1 1
R = .
i −i
Since f1 is an arbitrary complex constant, we may (cleverly) choose to write it as f1 = 21 (c1 + ic2 )
for arbitrary real constants c1 and c2 , and with this choice
−1 c
R F = 1 ,
c2
and so Y = P Z = P MZ F , i.e.,
⎡ ⎤ ⎡ (1+2i)x ⎤⎡ ⎤
−2 + 2i −2 − 2i 4 e 0 0 f1
Y = ⎣−1 + 2i −1 − 2i 3⎦ ⎣ 0 e(1−2i)x 0 ⎦ ⎣f1 ⎦ .
1 1 1 0 0 e5x c3
Now
⎡ ⎤ ⎡ ⎤
−2 + 2i −2 + 2i
e(1+2i)x ⎣−1 + 2i ⎦ = ex (cos(2x) + i sin(2x)) ⎣−1 + 2i ⎦
1 1
⎡ x ⎤ ⎡ x ⎤
e (−2 cos(2x) − 2 sin(2x)) e (2 cos(2x) − 2 sin(2x))
= ⎣ ex (− cos(2x) − 2 sin(2x)) ⎦ + i ⎣ ex (2 cos(2x) − sin(2x)) ⎦
ex cos(2x) ex sin(2x)
and of course ⎡ ⎤ ⎡ 5x ⎤
4 4e
5x ⎣ ⎦
e ⎣
3 = 3e5x ⎦ ,
1 e5x
so, replacing the relevant columns of P MZ , we find
⎡ ⎤⎡ ⎤
ex (−2 cos(2x) − 2 sin(2x)) ex (2 cos(2x) − 2 sin(2x)) 4e5x c1
Y = ⎣ ex (− cos(2x) − 2 sin(2x)) ex (2 cos(2x) − sin(2x)) 3e5x ⎦ ⎣c2 ⎦
ex cos(2x) ex sin(2x) e5x c3
⎡ ⎤
(−2c1 + 2c2 )e cos(2x) + (−2c1 − 2c2 )e sin(2x) + 4c3 e
x x 5x
the system Y = AY .
In Exercises 5and 6, solve
3 5 8
1. A = , det(λI − A) = λ2 − 8λ + 25, and Y0 = .
−2 5 13
3 4 3
2. A = , det(λI − A) = λ2 − 10λ + 29, and Y0 = .
−2 7 5
5 13 2
3. A = , det(λI − A) = λ2 − 14λ + 58, and Y0 = .
−1 9 1
7 17 5
4. A = , det(λI − A) = λ2 − 18λ + 145, and Y0 = .
−4 11 2
⎡ ⎤
37 10 20
5. A = ⎣−59 −9 −24⎦, det(λI − A) = (λ2 − 4λ + 29)(λ − 3).
−33 −12 −21
⎡ ⎤
−4 −42 15
6. A = ⎣ 4 25 −10⎦, det(λI − A) = (λ2 − 6λ + 13)(λ − 2).
6 32 −13
Y = AY + G .
2.3. INHOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 47
Y = (P J P −1 )Y + G
Y = P J (P −1 Y ) + G
P −1 Y = J (P −1 Y ) + P −1 G
(P −1 Y ) = J (P −1 Y ) + P −1 G .
Z = J Z + H
Y = PZ
is the solution to our original system.
Again, the key to this method is to be able to perform Step 2, and again this is straightforward.
Within each Jordan block, we solve from the bottom up. Let us focus our attention on a single k-by-k
block. The equation for the last function zk in that block is an inhomogeneous first-order differential
equation involving only zk , and we go ahead and solve it.The equation for the next to the last function
zk−1 in that block is an inhomogeneous first-order differential equation involving only zk−1 and zk .
We substitute in our solution for zk to obtain an inhomogeneous first-order differential equation for
zk−1 involving only zk−1 , and we go ahead and solve it, etc.
In principle, this is the method we use. In practice, using this method directly is solving each
system “by hand,” and instead we choose to “automate” this procedure. This leads us to the following
method. In order to develop this method we must begin with some preliminaries.
For a fixed matrix A, we say that the inhomogeneous system Y = AY + G(x) has associated
homogeneous system Y = AY . By our previous work, we know how to find the general solution of
Y = AY . First we shall see that, in order to find the general solution of Y = AY + G(x), it suffices
to find a single solution of that system.
Lemma 3.1. Let Yi be any solution of Y = AY + G(x). If Yh is any solution of the associated ho-
mogeneous system Y = AY , then Yh + Yi is also a solution of Y = AY + G(x), and every solution of
Y = AY + G(x) is of this form.
Consequently, the general solution of Y = AY + G(x) is given by Y = YH + Yi , where YH denotes
the general solution of Y = AY .
48 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
Proof. First we check that Y = Yh + Yi is a solution of Y = AY + G(x). We simply compute
as claimed.
Now we check that every solution Y of Y = AY + G(x) is of this form. So let Y be any
solution of this inhomogeneous system. We can certainly write Y = (Y − Yi ) + Yi = Yh + Yi where
Yh = Y − Yi . We need to show that Yh defined in this way is indeed a solution of Y = AY . Again
we compute
as claimed. 2
(It is common to call Yi a particular solution of the inhomogeneous system.)
Let us now recall our work from Section 2.1, and keep our previous notation. The homoge-
neous system Y = AY has general solution YH = P MZ C where C is a vector of arbitrary constants.
Let us set NY = NY (x) = P MZ (x) for convenience, so YH = NY C. Then YH = (NY C) = NY C,
and then, substituting in the equation Y = AY , we obtain the equation NY C = ANY C. Since this
equation must hold for any C, we conclude that
NY = ANY .
We use this fact to write down a solution to Y = AY + G. We will verify by direct computation
that the function we write down is indeed a solution. This verification is not a difficult one, but
nevertheless it is a fair question to ask how we came up with this function. Actually, it can be derived
in a very natural way, but the explanation for this involves the matrix exponential and so we defer it
until Section 2.4. Nevertheless, once we have this solution (no matter how we came up with it) we
are certainly free to use it.
It is convenient to introduce the following nonstandard notation. For a vector H (x), we
let
0 H (x)dx denote an arbitrary but fixed antiderivative of H (x). In other words, in obtaining
0 H (x)dx, we simply ignore the constants of integration. This is legitimate for our purposes, as by
Lemma 3.1 we only need to find a single solution to an inhomogeneous system, and it doesn’t matter
which one we find—any one will do. (Otherwise said, we can “absorb” the constants of integration
into the general solution of the associated homogeneous system.)
as claimed. 2
We now do a variety of examples: a 2-by-2 diagonalizable system, a 2-by-2 nondiagonalizable
system, a 3-by-3 diagonalizable system, and a 2-by-2 system in which the characteristic polynomial
has complex roots. In all these examples, when it comes to finding NY−1 , it is convenient to use the
fact that NY−1 = (P MZ )−1 = MZ−1 P −1 .
and NY = P MZ . Then
−3x −2x
−1 e 0 1 −1 30ex 6e − 12e−x
NY G = (1/5) = .
0 e2x −2 7 60e2x −12e3x + 84e4x
Then
−3e−2x + 12e2x
NY−1 G =
0 −4e3x + 21e4x
50 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
and
7 1 e3x 0 −3e−2x + 12e2x
Yi = NY NY−1 G
=
2 1 0 e−2x −4e3x + 21e4x
0x
−25e + 105e2x
= .
−10ex + 45e2x
and NY = P MZ . Then
−1 −2x 1 −x 0 −1 60e3x −18e3x − 60xex + 36xe3x
NY G = e (1/4) = .
0 1 4 −2 72e5x 60ex − 36e3x
Then
60ex − 60xex − 10e3x + 12xe3x
NY−1 G =
0 60ex − 12e3x
and
−2 1 2x 1 x 60ex − 60xex − 10e3x + 12xe3x
Yi = NY NY−1 G = e
−4 0 0 1 60ex − 12e3x
0 3x
−60e + 8e5x
= .
−240e3x + 40e5x
and NY = P MZ . Then
⎡ x ⎤⎡ ⎤⎡ x ⎤ ⎡ ⎤
e 0 0 0 1 1 e 12e4x + 20e5x
NY−1 G = ⎣ 0 1 0 ⎦ ⎣ 1 −2 −1⎦ ⎣12e3x ⎦ = ⎣ ex − 24e3x − 20e4x ⎦ .
0 0 e−2x −1 1 1 20e4x −e−x + 12ex + 20e2x
Then ⎡ ⎤
3e4x + 4e5x
NY−1 G = ⎣ ex − 8e3x − 5e4x ⎦
0
e−x + 12ex + 10e2x
and
⎡ ⎤ ⎡ −x ⎤⎡ ⎤
1 0 −1 e 0 0 3e4x + 4e5x
Yi = NY NY−1 G = ⎣0 −1 −1⎦ ⎣ 0 1 0 ⎦ ⎣ ex − 8e3x − 5e4x ⎦
0
1 1 1 0 0 e2x e−x + 12ex + 10e2x
⎡ x ⎤
−e − 9e3x − 6e4x
= ⎣−2ex − 4e3x − 5e4x ⎦ .
2ex + 7e3x + 9e4x
and NY = P MZ . Then
e−(3+4i)x 0 1 1 + 4i 200
NY−1 G = (1/(8i))
0 e−(3−4i)x 1 1 − 4i 160ex
−25e(−3−4i)x + 20(4 − i)e(−2−4i)x
= .
25e(−3+4i)x + 20(4 + i)e(−2+4i)x
52 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
Then
(4 + 3i)e(−3−4i)x + (−4 + 18i)e(−2−4i)x
NY−1 G =
0 (4 − 3i)e(−3+4i)x + (−4 − 18i)e(−2+4i)x
and
Yi = NY NY−1 G
0
−1 + 4i −1 − 4i e(3+4i)x 0 (4 + 3i)e(−3−4i)x + (−4 + 18i)e(−2−4i)x
=
1 1 0 e(3−4i)x (4 − 3i)e(−3+4i)x + (−4 − 18i)e(−2+4i)x
−1 + 4i −1 − 4i (4 + 3i) + (−4 + 18i)ex
=
1 1 (4 − 3i) + (−4 − 18i)ex
−32 − 136e x
= .
8 − 8ex
(Note that in this last example we could do arithmetic with complex numbers directly, i.e.,
without having to convert complex exponentials into real terms.)
Once we have done this work, it is straightforward to solve initial value problems. We do a
single example that illustrates this.
Definition 4.1. Let T be a square matrix. Then the matrix exponential eT is defined by
1 2 1 1
eT = I + T + T + T3 + T4 + ... .
2! 3! 4!
(For this definition to make sense we need to know that this series always converges, and it
does.)
Recall that the differential equation y = ay has the solution y = ceax . The situation for
Y = AY is very analogous. (Note that we use rather than C to denote a vector of constants for
reasons that will become clear a little later. Note that is on the right in Theorem 4.2 below, a
consequence of the fact that matrix multiplication is not commutative.)
Theorem 4.2.
Y = AY
is given by
Y = eAx
has solution
Y = eAx Y0 .
Proof. (Outline) (1) We first compute eAx . In order to do so, note that (Ax)2 = (Ax)(Ax) =
(AA)(xx) = A2 x 2 as matrix multiplication commutes with scalar multiplication, and (Ax)3 =
(Ax)2 (Ax) = (A2 x 2 )(Ax) = (A2 A)(x 2 x) = A3 x 3 , and similarly, (Ax)k = Ak x k for any k. Then,
substituting in Definition 4.1, we have that
1 2 2 1 1
Y = eAx = (I + Ax + A x + A3 x 3 + A4 x 4 + . . .) .
2! 3! 4!
2.4. THE MATRIX EXPONENTIAL 55
To find Y , we may differentiate this series term-by-term. (This claim requires proof, but we shall
not give it here.) Remembering that A and are constant matrices, we see that
1 2 1 1
Y = (A + A (2x) + A3 (3x 2 ) + A4 (4x 3 ) + . . .)
2! 3! 4!
1 1
= (A + A2 x + A3 x 2 + A4 x 3 + . . .)
2! 3!
1 2 2 1
= A(I + Ax + A x + A3 x 3 + . . .)
2! 3!
= A(eAx ) = AY
as claimed.
(2) By (1) we know that Y = AY has solution Y = eAx . We use the initial condition to
solve for . Setting x = 0, we have:
Y0 = Y (0) = eA0 = e0 = I =
(where e0 means the exponential of the zero matrix, and the value of this is the identity matrix I , as
is apparent from Definition 4.1), so = Y0 and Y = eAx = eAx Y0 . 2
In the remainder of this section we shall see how to translate the theoretical solution of
Y = AY given by Theorem 4.2 into a practical one. To keep our notation simple, we will stick to
2-by-2 or 3-by-3 cases, but the principle is the same regardless of the size of the matrix.
One case is relatively easy.
To do so we directly apply Theorem 4.2 and Lemma 4.3. The solution is given by
3x
y1 e 0 γ1 γ1 e3x
=Y =e =
Jx
= .
y2 0 e−2x γ2 γ2 e−2x
5 −7
Now suppose we want to find the general solution of Y = AY where A = . We may
2 −4
still apply Theorem 4.2 to conclude that the solution is Y = eAx . We again try to calculate eAx .
Now we find
5 −7 11 −7 41 −49
A= , A2 = , A3 = , ...
2 −4 2 2 14 −22
2.4. THE MATRIX EXPONENTIAL 57
so
1 0 5 −7 1 11 −7 2 1 41 −49 3
e Ax
= + x+ x + x + ... ,
0 1 2 −4 2! 2 2 3! 14 −22
which looks like a hopeless mess. But, in fact, the situation is not so hard!
S = P T P −1
S k = P T k P −1 for every k
and
eS = P eT P −1 .
S 2 = SS = (P T P −1 )(P T P −1 ) = P T (P −1 P )T P −1 = P T I T P −1
= P T T P −1 = P T 2 P −1 ,
S 3 = S 2 S = (P T 2 P −1 )(P T P −1 ) = P T 2 (P −1 P )T P −1 = P T 2 I T P −1
= P T 2 T P −1 = P T 3 P −1 ,
S 4 = S 3 S = (P T 3 P −1 )(P T P −1 ) = P T 3 (P −1 P )T P −1 = P T 3 I T P −1
= P T 3 T P −1 = P T 4 P −1 ,
etc.
Then
1 2 1 1
eS = I + S + S + S3 + S4 + . . .
2! 3! 4!
1 1 1
= P I P −1 + P T P −1 + P T 2 P −1 + P T 3 P −1 + P T 4 P −1 + . . .
2! 3! 4!
1 1 1
= P (I + T + T 2 + T 3 + T 4 + . . .)P −1
2! 3! 4!
= P eT P −1
as claimed. 2
58 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
With this in hand let us return to our problem.
Example 4.6. (Compare Example 1.2.) We wish to find the general solution of Y = AY where
5 −7
A= .
2 −4
Then
eAx = P eJ x P −1
−1
7 1 e3x 0 7 1
=
2 1 0 e−2x 2 1
7 3x 2 −2x
5e − 5e − 75 e3x + 75 e−2x
=
2 −2x
5e − 5e
2 3x
− 25 e3x + 75 e−2x
and
γ1
Y =e Ax
=e Ax
γ2
( 75 γ1 − 75 γ2 )e3x + (− 25 γ1 + 75 γ2 )e−2x
= .
( 25 γ1 − 25 γ2 )e3x + (− 25 γ1 + 75 γ2 )e−2x
Example 4.7. (Compare Example 1.3.) We wish to find the general solution of Y = AY where
⎡ ⎤
2 −3 −3
A = ⎣ 2 −2 −2⎦ .
−2 1 1
eAx = P eJ x P −1
⎡ ⎤ ⎡ −x ⎤⎡ ⎤−1
1 0 −1 e 0 0 1 0 −1
= ⎣0 −1 −1⎦ ⎣ 0 1 0 ⎦ ⎣0 −1 −1⎦
1 1 1 0 0 e2x 1 1 1
⎡ −x −x ⎤
e 2x e −e 2x e −e 2x
and
⎡ ⎤
γ1
Y = eAx = eAx ⎣γ2 ⎦
γ3
⎡ ⎤
(γ2 + γ3 )e−x + (γ1 − γ2 − γ3 )e2x
= ⎣ (−γ1 + 2γ2 + γ3 ) + (γ1 − γ2 − γ3 )e2x ⎦ .
−x
(γ2 + γ3 )e + (γ1 − 2γ2 − γ3 ) + (−γ1 + γ2 + γ3 )e 2x
⎡ ⎤
1
Now suppose we want to solve the initial value problem Y = AY , Y (0) = ⎣0⎦. Then
0
Y = eAx Y (0)
⎡ ⎤⎡ ⎤
e2x e−x − e2x e−x − e2x 1
= ⎣−1 + e 2x 2 − e2x 1 − e2x ⎦ ⎣0⎦
1 − e2x e−x − 2 + e2x e−x − 1 + e2x 0
⎡ ⎤
e2x
= ⎣−1 + e2x ⎦ .
1 − e2x
Remark 4.8. Let us compare the results of our method here with that of our previous method. In
the case of Example 4.6, our previous method gives the solution
3x
e 0
Y =P C
0 e−2x
= P eJ x C
60 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
3 0
where J = ,
0 −2
while our method here gives
Y = P eJ x P −1 .
But note that these answers are really the same! For P −1 is a constant matrix, so if is a vector of
arbitrary constants, then so is P −1 , and we simply set C = P −1 .
Similarly, in the case of Example 4.7, our previous method gives the solution
⎡ ⎤
e−x 0 0
Y =P⎣ 0 1 0 ⎦C
0 0 e2x
= P eJ x C
⎡ ⎤
−1 0 0
where J = ⎣ 0 0 0⎦,
0 0 2
while our method here gives
Y = P eJ x P −1
and again, setting C = P −1 , we see that these answers are the same.
So the point here is not that the matrix exponential enables us to solve new problems, but
rather that it gives a new viewpoint about the solutions that we have already obtained.
While these two methods are in principle the same, we may ask which is preferable in
practice. In this regard we see that our earlier method is better, as the use of the matrix exponential
requires us to find P −1 , which may be a considerable amount of work. However, this advantage
is (partially) negated if we wish to solve initial value problems, as the matrix exponential method
immediately gives the unknown constants , as = Y (0), while in the former method we must
solve a linear system to obtain the unknown constants C.
Now let us consider the nondiagonalizable case. Suppose Z = J Z where J is a matrix con-
sisting of a single Jordan block. Then by Theorem 4.2 this has the solution Z = eJ x . On the other
2.4. THE MATRIX EXPONENTIAL 61
hand, in Theorem 1.1 we already saw that this system has solution Z = MZ C. In this case, we simply
have C = , so we must have eJ x = MZ . Let us see that this is true by computing eJ x directly.
Then
⎡ ⎤
1 x x 2 /2! x 3 /3! · · · x k−1 /(k − 1)!
⎢0 1 x x 2 /2! · · · x k−2 /(k − 2)!⎥
⎢ ⎥
⎢ 1 x · · · x k−3 /(k − 3)!⎥
⎢ ⎥
eJ x = eax ⎢ . . ⎥.
⎢ . . .
. ⎥
⎢ ⎥
⎣ x ⎦
1
1 2 1 1 1 1 1
m13 = x + ax 3 + ( a 2 x 4 ) + ( a 3 x 5 ) + . . .
2! 2! 2! 2! 2! 3!
(as 6/4! = 1/4 = (1/2!)(1/2!) and 10/5! = 1/12 = (1/2!)(1/3!), etc.)
1 2 1 1
= x (1 + ax + (ax)2 + (ax)3 + . . .)
2! 2! 3!
1
= x 2 eax ,
2!
so
⎡ 1 2 ax ⎤ ⎡ ⎤
eax xeax 2! x e 1 x x 2 /2!
eJ x =⎣ 0 eax xeax ⎦ = eax ⎣0 1 x ⎦ ,
0 0 eax 0 0 1
Example 4.10. (Compare Examples 1.6 and 1.10.) Consider the system
0 1
Y = AY where A = .
−4 4
3
Also, consider the initial value problem Y = AY , Y (0) = .
−8
We saw in Example 2.12 in Chapter 1 that A = P J P −1 with
−2 1 2 1
P = and J = .
−4 0 0 2
Then
eAx = P eJ x P −1
−1
−2 1 e2x xe2x −2 1
=
−4 0 0 e2x −4 0
(1 − 2x)e2x xe2x
= ,
−4xe2x (1 + 2x)e2x
64 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
and so
Y = eAx
Ax γ1
=e
γ2
γ1 e2x + (−2γ1 + γ2 )xe2x
= .
γ2 e2x + (−4γ1 + 2γ2 )xe2x
Y = eAx Y0
3
=e Ax
−8
2x
3e − 14xe2x
= .
8e2x − 28xe2x
Example 4.11. (Compare Examples 1.7 and 1.11.) Consider the system
⎡ ⎤
2 1 1
Y = AY where A = ⎣ 2 1 −2⎦ .
−1 0 2
⎡ ⎤
8
Also, consider the initial value problem Y = AY , Y (0) = ⎣32⎦.
5
We saw in Example 2.25 in Chapter 1 that A = P J P −1 with
⎡ ⎤ ⎡ ⎤
1 0 −5 −1 1 0
P = ⎣−2 0 6 ⎦ and J = ⎣ 0 −1 0⎦ .
−1 1 1 0 0 3
Then
eAx = P eJ x P −1
⎡ ⎤ ⎡ −x ⎤⎡ ⎤−1
1 0 −5 e xe−x 0 1 0 −5
⎣
= −2 0 6 ⎦ ⎣ 0 e−x 0 ⎦ ⎣−2 0 6 ⎦
−1 1 1 0 0 e3x −1 1 1
2.4. THE MATRIX EXPONENTIAL 65
⎡3 −x + −x 5 −x ⎤
8e + 1
2 xe
5 3x
8e − 16 e − 41 xe−x + 165 3x
e xe−x
⎢ 3 ⎥
=⎢ −x −x
⎣− 4 e − xe +
3 3x
4e
5 −x
8e
−x
+ 2 xe + 8 e
1 3 3x
−2xe−x ⎥
⎦
1 −x
8e − 21 xe−x − 1 3x
8e
1 −x
16 e + 41 xe−x − 16
1 3x
e e−x − xe−x
and so
Y = eAx
⎡ ⎤
γ1
Ax ⎣ ⎦
=e γ2
γ3
⎡ ⎤
( 38 γ1 − 16
5
γ2 )e−x + ( 21 γ1 − 41 γ2 + γ3 )xe−x + ( 58 γ1 + 16
5
γ2 )e3x
⎢ ⎥
=⎢⎣ (− 43 γ1 + 58 γ2 )e−x + (−γ1 + 21 γ2 − 2γ3 )xe−x + ( 43 γ1 + 83 γ2 )e3x ⎥ .
⎦
( 18 γ1 + 1
16 γ2 + γ3 )e−x + (− 21 γ1 + 41 γ2 − γ3 )xe−x + (− 18 γ1 − 1
16 γ2 )e
3x
Y = eAx Y0
⎡ ⎤
8
= eAx ⎣32⎦
5
⎡ ⎤
−7e−x + xe−x + 15e3x
= ⎣14e−x − 2xe−x + 18e3x ⎦ .
8e−x − xe−x − 3e3x
Now we solve Y = AY in an example where the matrix A has complex eigenvalues. As you
will see, our method is exactly the same.
Remark 4.13. Our procedure in this section is essentially that of Remark 1.12. (Compare Exam-
ple 4.10 with Example 1.13.)
Remark 4.14. As we have seen, for a matrix J in JCF, eJ x = MZ , in the notation of Section 2.1.
But also, in the notation of Section 2.1, if A = P J P −1 , then eAx = P eJ x P −1 = P MZ P −1 = MY .
2.4. THE MATRIX EXPONENTIAL 67
Remark 4.15. Now let us see how to use the matrix exponential to solve an inhomogeneous system
Y = AY + G(x). Since we already know how to solve homogeneous systems, we need only, by
Lemma 3.1, find a (single) particular solution Yi of this inhomogeneous
system, and that is what
we do. We shall again use our notation from Section 2.3, that 0 H (x)dx denotes an arbitrary (but
fixed) antiderivative of H (x).
Y = AY + G(x)
Y − AY = G(x)
with solution
−Ax
e Yi = e−Ax G(x)
0
Yi = e Ax
e−Ax G(x) .
0
Let us compare this with the solution we found in Theorem 3.2. By Remark 4.14, we can
rewrite this solution as Yi = MY 0 MY−1 G(x). This is almost, but not quite, what we had in Theo-
rem 3.2. There we had the solution Yi = NY 0 NY−1 G(x), where NY = P MZ . But these solutions
are the same, as MY = P MZ P −1 = NY P −1 . Then MY−1 = P MZ−1 P −1 and NY−1 = MZ−1 P −1 , so
MY−1 = P NY−1 . Substituting, we find
Yi = MY MY−1 G(x)
0
= NY P −1
P NY−1 G(x) ,
0
and, since P is a constant matrix, we may bring it outside the integral to obtain
Yi = NY P P NY−1 G(x)
−1
0
−1
= NY NY G(x)
0
as claimed.
68 CHAPTER 2. SOLVING SYSTEMS OF LINEAR DIFFERENTIAL EQUATIONS
Remark 4.16. In applying this method we must compute MZ−1 = (eJ x )−1 = e−J x = eJ (−x) ,
and, as an aid to calculation, it is convenient to make the following observation. Suppose, for
simplicity, that J consists of a single Jordan block. Then we compute: in the 1-by-1 case,
−1
ax −1 −ax 1 x −ax 1 −x
e = e ; in the 2-by-2 case, e ax =e ; in the 3-by-3 case,
0 1 0 1
⎛ ⎡ ⎤⎞−1 ⎡ ⎤ ⎡ ⎤
1 x x 2 /2! 1 −x (−x)2 /2! 1 −x x 2 /2!
⎝eax ⎣0 1 x ⎦⎠ = e−ax ⎣0 1 −x ⎦ = e−ax ⎣0 1 −x ⎦, etc.
0 0 1 0 0 1 0 0 1
(b) Use part (a) to solve the initial value problem Y = AY , Y (0) = Y0 .
Exercises 1–24: In Exercise n, for 1 ≤ n ≤ 20, the matrix A and the initial vector Y0 are the
same as in Exercise n of Section 2.1. In Exercise n, for 21 ≤ n ≤ 24, the matrix A and the initial
vector Y0 are the same as in Exercise n − 20 of Section 2.2.
69
APPENDIX A
Background Results
A.1 BASES, COORDINATES, AND MATRICES
In this section of the Appendix, we review the basic facts on bases for vector spaces and on coordinates
for vectors and matrices for linear transformations.Then we use these to (re)prove some of the results
in Chapter 1.
First we see how to represent vectors, once we have chosen a basis.
Theorem 1.1. Let V be a vector space and let B = {v1 , . . . , vn } be a basis of V . Then any vector v in
V can be written as v = c1 v1 + . . . + cn vn in a unique way.
Definition 1.2. Let V be a vector space and let B = {v1 , . . . , vn } be a basis of V . Let v be a vector
in V and write v = c1 v1 + . . . + cn vn . Then the vector
⎡ ⎤
c1
⎢ .. ⎥
[v]B = ⎣ . ⎦
cn
Remark 1.3. In particular, we may take V = Cn and consider the standard basis E = {e1 , . . . , en }
where
⎡ ⎤
0
⎢ .. ⎥
⎢.⎥
⎢ ⎥
⎢0⎥
ei = ⎢
⎢1⎥ ,
⎥
⎢ ⎥
⎢0⎥
⎣ ⎦
..
.
Next we see how to represent linear transformations, once we have chosen a basis.
Theorem 1.4. Let V be a vector space and let B = {v1 , . . . , vn } be a basis of V . Let T : V −→ V be
a linear transformation. Then there is a unique matrix [T ]B such that, for any vector v in V ,
[T (v)]B = [T ]B [v]B .
Definition 1.5. Let V be a vector space and let B = {v1 , . . . , vn } be a basis of V . Let T : V −→ V
be a linear transformation. Let [T ]B be the matrix defined in Theorem 1.4. Then [T ]B is the matrix
of the linear transformation T in the basis B .
A.1. BASES, COORDINATES, AND MATRICES 71
Remark 1.6. In particular, we may take V =and Cn
consider
the standard
basis E = {e1 , . . . , en }.
Let A be an n-by-n square matrix and write A = a1 a2 . . . an . If TA is the linear transfor-
mation given by TA (v) = Av, then
[TA ]E = [TA (e1 )]E [TA (e2 )]E . . . [TA (en )]E
= [Ae1 ]E [Ae2 ]E . . . [Aen ]E
= [a1 ]E [a2 ]E . . . [an ]E
= a1 a2 . . . an
=A.
(In other words, the linear transformation given by multiplication by a matrix “looks like” that same
matrix in the standard basis.)
What is essential to us is the ability to compare the situation in different bases. To that end,
we have the following theorem.
Theorem 1.7. Let V be a vector space, and let B = {v1 , . . . , vn } and C = {w1 , . . . , wn } be two bases
of V . Let PC ←B be the matrix
PC ←B = [v1 ]C [v2 ]C . . . [vn ]C .
[T ]C = PC ←B [T ]B PB←C
= PC ←B [T ]B (PC ←B )−1
= (PB←C )−1 [T ]B PB←C .
72 APPENDIX A. BACKGROUND RESULTS
Again, this theorem leads to a definition.
Definition 1.8. The matrix PC ←B is the change-of-basis matrix from the basis B to the basis C .
Corollary 1.9. Let V = Cn , let E be the standard basis of V , and let B = {v1 , . . . , vn } be any basis of
V . Let A be any n-by-n square matrix. Then
A = P BP −1
where
P = v1 v2 . . . vn and B = [TA ]B .
With this in hand we now present new proofs of Theorems 1.14 and 2.11 in Chapter 1, and
a proof of Lemma 1.7 in Chapter 1. For convenience, we restate these results.
Theorem 1.10. Let A be an n-by-n matrix over the complex numbers. Then A is diagonalizable if
and only if, for each eigenvalue a of A, geom-mult(a) = alg-mult(a). In that case, A = P J P −1 where
J is a diagonal matrix whose entries are the eigenvalues of A, each appearing according to its algebraic
multiplicity, and P is a matrix whose columns are eigenvectors forming bases for the associated eigenspaces.
A.1. BASES, COORDINATES, AND MATRICES 73
Proof. First suppose that for each eigenvalue a of A, geom-mult(a) = alg-mult(a). In the notation of
the proof of Theorem 1.14 in Chapter 1, B = {v1 , . . . , vn } is a basis of Cn . Then, by Corollary 1.9,
A = P [TA ]B P −1 . But B is a basis of eigenvectors, so for each i, TA (vi ) = Avi = ai vi = 0v1 +
. . . + 0vi−1 + ai vi + 0vi+1 + . . . + 0vn . Then
⎡ ⎤
0
⎢ .. ⎥
⎢.⎥
⎢ ⎥
[TA (vi )]B = ⎢
⎢ai ⎥
⎥
⎢ .. ⎥
⎣.⎦
0
with ai in the i th position and 0 elsewhere. But [TA (vi )]B is the i th column of the matrix [TA ]B , so
we see that J = [TA ]B is a diagonal matrix.
Conversely, if J = [TA ]B is a diagonal matrix, then the same computation shows that Avi =
ai vi , so for each i, vi is an eigenvector of A with associated eigenvalue ai . 2
Theorem 1.11. Let A be a k-by-k matrix and suppose that Ck has a basis {v1 , . . . , vk } consisting of a
single chain of generalized eigenvectors of length k associated to an eigenvalue a. Then
A = P J P −1
where ⎡ ⎤
a 1
⎢ a 1 ⎥
⎢ ⎥
⎢ a 1 ⎥
⎢ ⎥
J =⎢ .. .. ⎥
⎢ . . ⎥
⎢ ⎥
⎣ a 1⎦
a
is a matrix consisting of a single Jordan block and
P = v1 v2 . . . vk
Proof. Let B = {v1 , . . . , vk }. Then, by Corollary 1.9, A = P [TA ]B P −1 . Now the i th column of
[TA ]B is [TA (vi )]B = [Avi ]B . By the definition of a chain,
Avi = (A − aI + aI )vi
= (A − aI )vi + aI vi
= vi−1 + avi for i > 1, = avi for i = 1 ,
74 APPENDIX A. BACKGROUND RESULTS
so for i > 1
⎡ ⎤
0
⎢ .. ⎥
⎢.⎥
⎢ ⎥
⎢1⎥
[Avi ]B = ⎢
⎢a ⎥
⎥
⎢ ⎥
⎢0 ⎥
⎣ ⎦
..
.
with 1 in the (i − 1)st position, a in the i th position, and 0 elsewhere, and [Av1 ]B is similar, except
that a is in the 1st position (there is no entry of 1), and every other entry is 0. Assembling these
vectors, we see that the matrix [TA ]B = J has the form of a single k-by-k Jordan block with diagonal
entries equal to a. 2
1 ≤ geom-mult(a) ≤ alg-mult(a) .
Proof. By the definition of an eigenvalue, there is at least one eigenvector v with eigenvalue a, and
so Ea contains the nonzero vector v, and hence dim(Ea ) ≥ 1.
Now suppose that a has geometric multiplicity k, and let {v1 , . . . , vk } be a basis for the
eigenspace Ea . Extend this basis to a basis B = {v1 , . . . , vk , vk+1 , . . . , vn } of Cn . Let B = [TA ]B .
Then
B = b1 b2 . . . bn
with bi = [TA (vi )]B . But for i between 1 and k, TA (vi ) = Avi = avi , so
⎡ ⎤
0
⎢ .. ⎥
⎢.⎥
⎢ ⎥
bi = [avi ]B = ⎢
⎢a ⎥
⎥
⎢ .. ⎥
⎣.⎦
0
with a in the i th position and 0 elsewhere. (For i > k, we do not know what bi is.)
Now we compute the characteristic polynomial of B, det(λI − B). From our computation of
B, we see that, for i between 1 and k, the i th column of (λI − B) is
A.2. PROPERTIES OF THE COMPLEX EXPONENTIAL 75
⎡ ⎤
0
⎢ .. ⎥
⎢ . ⎥
⎢ ⎥
⎢λ − a ⎥
⎢ ⎥
⎢ .. ⎥
⎣ . ⎦
0
with λ − a in the i th position and 0 elsewhere. (For i > k, we do not know what the i th column of
this matrix is.)
To compute the characteristic polynomial, i.e., the determinant of this matrix, we successively
expand by minors of the 1st , 2nd , . . ., k th columns. Each of these gives a factor of (λ − a), so we see
that det(λI − B) = (λ − a)k q(λ) for some (unknown) polynomial q(λ).
We have computed the characteristic polynomial of B, but what we need to know is the
characteristic polynomial of A. But these are equal, as we see from the following computation
(which uses the fact that scalar multiplication commutes with matrix multiplication, and properties
of determinants):
det(λI − A) = det(λI − P BP −1 ) = det(λ(P I P −1 ) − P BP −1 )
= det(P (λI )P −1 − P BP −1 ) = det(P (λI − B)P −1 )
= det(P ) det(λI − B) det(P −1 ) = det(P ) det(λI − B)(1/ det(P ))
= det(λI − B) .
Thus, det(λI − A), the characteristic polynomial of A, is divisible by (λ − a)k (and perhaps
by a higher power of (λ − a), and perhaps not, as we do not know anything about the polynomial
q(λ)), so alg-mult(a) ≥ k = geom-mult(a), as claimed. 2
First we note that this definition indeed makes sense, as this power series converges for every
complex number z. Now for the properties we wish to prove. Note that properties (2) and (3) are
direct generalizations of the situation for the real exponential function.
eiθ = 1 + iθ + (iθ)2 /2! + (iθ )3 /3! + (iθ )4 /4! + (iθ )5 /5! + (iθ )6 /6! + (iθ )7 /7! + . . .
= 1 + iθ − θ 2 /2! − iθ 3 /3! + θ 4 /4! + iθ 5 /5! − θ 6 /6! − iθ 7 /7! + . . . .
We now rearrange the terms, gathering the terms that do not involve i together and gathering
the terms that do involve i together. (That we can do so requires proof, but we shall not give that
proof here.) We obtain:
But we recognize the power series inside the first set of parentheses as the power series for
cos(θ), and the power series inside the second set of parentheses as the power series for sin(θ),
completing the proof.
(2) We prove a more general formula. Consider the function ec+dz , where c and d are arbitrary
constants. To differentiate this function, we substitute in the power series and differentiate term-
by-term (again a procedure that requires justification, but whose justification we again skip). Using
the chain rule to take the derivative of each term, we obtain
(5) This follows directly from (1) and (4). Let z = s + it, so z = s − it. We compute:
2
78
79
APPENDIX B
Answers to Odd-Numbered
Exercises
Chapter 1
−1
−7 4 3 0 −7 4
1. A = .
9 −5 0 5 9 −5
−1
−21 1 3 1 −21 1
3. A = .
−49 0 0 3 −49 0
−1
−5 1 7 1 −5 1
5. A = .
−25 0 0 7 −25 0
⎡ ⎤⎡ ⎤⎡ ⎤−1
0 2 0 −1 0 0 0 2 0
7. A = ⎣1 1 −3⎦ ⎣ 0 1 0⎦ ⎣1 1 −3⎦ .
1 1 1 0 0 3 1 1 1
⎡ ⎤⎡ ⎤⎡ ⎤−1
−1 −2 −2 −3 0 0 −1 −2 −2
9. A = ⎣ 1 0 −1⎦ ⎣ 0 −3 0⎦ ⎣ 1 0 −1⎦ .
0 1 1 0 0 1 0 1 1
⎡ ⎤⎡ ⎤⎡ ⎤−1
−1 −2 −1 4 0 0 −1 −2 −1
11. A = ⎣ 0 1 1 ⎦ ⎣0 4 0⎦ ⎣ 0 1 1⎦ .
1 0 1 0 0 2 1 0 1
⎡ ⎤⎡ ⎤⎡ ⎤−1
−1 0 0 −2 1 0 −1 0 0
13. A = ⎣−1 0 1⎦ ⎣ 0 −2 0⎦ ⎣−1 0 1⎦ .
0 1 1 0 0 4 0 1 1
⎡ ⎤⎡ ⎤⎡ ⎤−1
−3 −1 −2 0 1 0 −3 −1 −2
15. A = ⎣−2 −1 −2⎦ ⎣0 0 0⎦ ⎣−2 −1 −2⎦ .
3 1 3 0 0 3 3 1 3
80 APPENDIX B. ANSWERS TO ODD-NUMBERED EXERCISES
⎡ ⎤⎡ ⎤⎡ ⎤−1
−2 1 1 1 1 0 −2 1 1
17. A = ⎣−10 0 2⎦ ⎣0 1 0⎦ ⎣−10 0 2⎦ .
−6 0 0 0 0 1 −6 0 0
⎡ ⎤⎡ ⎤⎡ ⎤−1
1 2 0 0 1 0 1 2 0
19. A = ⎣2 3 0⎦ ⎣0 0 1⎦ ⎣2 3 0⎦ .
1 3 1 0 0 0 1 3 1
Section 2.1
−7c1 e3x + 4c2 e5x −7e3x + 8e5x
1a. Y = . b. Y = .
9c1 e3x − 5c2 e5x 9e3x − 10e5x
3x
(−21c1 + c2 )e3x − 21c2 xe3x 41e + 21xe3x
3a. Y = . b. Y = .
−49c1 e3x − 49c2 xe3x 98e3x + 49xe3x
(−5c1 + c2 )e7x − 5c2 xe7x −10e7x − 25xe7x
5a. Y = . b. Y = .
−25c1 e7x − 25c2 xe7x −75e7x − 125xe7x
⎡ ⎤ ⎡ ⎤
2c2 ex 6ex
7a. Y = ⎣c1 e−x + c2 ex − 3c3 e3x ⎦. b. Y = ⎣2e−x + 3ex − 15e3x ⎦.
c1 e−x + c2 ex + c3 e3x 2e−x + 3ex + 5e3x
⎡ ⎤ ⎡ ⎤
(−c1 − 2c2 )e−3x − 2c3 ex 0
9a. Y = ⎣ c1 e−3x − c3 ex ⎦. b. Y = ⎣ 2e−3x ⎦.
c2 e−3x + c3 ex −e−3x
⎡ ⎤ ⎡ ⎤
(−c1 − 2c2 )e4x − c3 e2x 2e4x − 5e2x
11a. Y = ⎣ c2 e4x + c3 e2x ⎦. b. Y = ⎣−3e4x + 5e2x ⎦.
c1 e4x + c3 e2x 4e4x + 5e2x
⎡ ⎤ ⎡ ⎤
−c1 e−2x − c2 xe−2x −e−2x − 2xe−2x
13a. Y = ⎣−c1 e−2x − c2 xe−2x + c3 e4x ⎦. b. Y = ⎣−e−2x − 2xe−2x + 4e4x ⎦.
c2 e−2x + c3 e4x 2e−2x + 4e4x
⎡ ⎤ ⎡ ⎤
(−3c1 − c2 ) − 3c2 x − 2c3 e3x −9 − 9x + 10e3x
15a. Y = ⎣(−2c1 − c2 ) − 2c2 x − 2c3 e3x ⎦. b. Y = ⎣−7 − 6x + 10e3x ⎦.
(3c1 + c2 ) + 3c2 x + 3c3 e3x 9 + 9x − 15e3x
81
⎡ ⎤ ⎡ ⎤
(−2c1 + c2 + c3 − 2c2
)ex xex 3ex − 14xex
⎣
17a. Y = (−10c1 + 2c3 )e − 10c2 xex ⎦.
x b. Y = ⎣10ex − 70xex ⎦.
−6c1 ex − 6c2 xex 18ex − 42xex
⎡ ⎤ ⎡ ⎤
(c1 + 2c2 ) + (c2 + 2c3 )x + (c3 /2)x 2 6 + 5x + x 2
19a. Y = ⎣ (2c1 + 3c2 ) + (2c2 + 3c3 )x + c3 x 2 ⎦. b. Y = ⎣11 + 8x + 2x 2 ⎦.
(c1 + 3c2 + c3 ) + (c2 + 3c3 )x + (c3 /2)x 2 9 + 7x + x 2
Section 2.2
1a.
e4x (cos(3x) + 3 sin(3x)) e4x (−3 cos(3x) + sin(3x)) c1
Y =
2e4x cos(3x) 2e4x sin(3x) c2
(c1 − 3c2 )e cos(3x) + (3c1 + c2 )e sin(3x)
4x 4x
= .
2c1 e4x cos(3x) + 2c2 e4x sin(3x)
8e4x cos(3x) + 19e4x sin(3x)
b. Y = .
13e4x cos(3x) − sin(3x)
3a.
e7x (2 cos(3x) + 3 sin(3x)) e7x (−3 cos(3x) + 2 sin(3x)) c1
Y =
e7x cos(3x) e7x sin(3x) c2
(2c1 − 3c2 )e cos(3x) + (3c1 + 2c2 )e sin(3x)
7x 7x
= .
c1 e7x cos(3x) + c2 e7x sin(3x)
2e7x cos(3x) + 3e7x sin(3x)
b. Y = .
e7x cos(3x)
5.
⎡ 2x ⎤⎡ ⎤
e (− cos(5x) − sin(5x)) e2x (cos(5x) − sin(5x)) 0 c1
⎣
Y = e (−3 cos(5x) + 4 sin(5x)) e (−4 cos(5x) − 3 sin(5x)) −2e
2x 2x 3x ⎦ ⎣c2 ⎦
2x
3e cos(5x) 2x
3e sin(5x) e 3x c3
⎡ ⎤
(−c1 + c2 )e cos(5x) + (−c1 − c2 )e sin(5x)
2x 2x
= ⎣(−3c1 − 4c2 )e2x cos(5x) + (4c1 − 3c2 )e2x sin(5x) − 2c3 e3x ⎦ .
3c1 e2x cos(5x) + 3c2 e2x sin(5x) + c3 e3x
Section 2.3
10e8x − 168e4x
1. Yi = .
−12e8x + 213e4x
82 APPENDIX B. ANSWERS TO ODD-NUMBERED EXERCISES
−20e4x + 9e5x
3. Yi = .
−49e4x + 23e5x
−2e10x + e12x
5. Yi = .
−25e10x + 10e12x
⎡ ⎤
−1
7. Yi = ⎣−1 − 2e2x − 3e4x ⎦.
−1 + e2x + 2e4x
Section 2.4
−35e3x + 36e5x −28e3x + 28e5x
1a. eAx = .
45e3x − 45e5x 36e3x − 35e5x
(−35γ1 − 28γ2 )e3x + (36γ1 + 28γ2 )e5x
Y = .
(45γ1 + 36γ2 )e3x + (−45γ1 − 35γ2 )e5x
e3x − 21xe3x 9xe3x
3a. eAx = .
−49xe3x e3x + 21xe3x
γ1 e3x + (−21γ1 + 9γ2 )xe3x
Y = .
γ2 e3x + (−49γ1 + 21γ2 )xe3x
e7x − 5xe7x xe7x
5a. eAx = .
−25xe7x e7x + 5xe7x
γ1 e7x + (−5γ1 + γ2 )xe7x
Y = .
γ2 e7x + (−25γ1 + 5γ2 )xe7x
⎡ ⎤
ex 0 0
⎢ 1 −x 1 x 1 −x 3 −x
− 43 e3x ⎥
7a. eAx = ⎣− 2 e + 2 e 4e + 43 e3x 4e ⎦.
− 21 e−x + 21 ex 1 −x
4e − 41 e3x 3 −x
4e + 41 e3x
⎡ ⎤
γ1 ex
⎢ −x 3x ⎥
Y = ⎣ (− 2 γ1 + 4 γ2 + 4 γ3 )e + 2 γ1 e + ( 4 γ2 − 4 γ3 )e ⎦ .
1 1 3 1 x 3 3
(− 21 γ1 + 41 γ2 + 43 γ3 )e−x + 21 γ1 ex + (− 41 γ2 + 41 γ3 )e3x
83
⎡ ⎤
−e−3x + 2ex −2e−3x + 2ex −4e−3x + 4ex
9a. eAx = ⎣ −e−3x + ex ex −2e−3x + 2ex ⎦ .
e−3x − ex e−3x − ex 3e−3x − 2ex
⎡ ⎤
(−γ1 − 2γ2 − 4γ3 )e−3x + (2γ1 + 2γ2 + 4γ3 )ex
Y = ⎣ (−γ1 − 2γ3 )e−3x + (γ1 + γ2 + 2γ3 )ex ⎦.
(γ1 + γ2 + 2γ3 )e −3x + (−γ1 − γ2 − 2γ3 )e x
⎡ 1 2x ⎤
2e − 2e e4x − e2x 2e − 2e
3 4x 1 2x 1 4x
⎢ 1 ⎥
11a. eAx = ⎢− e4x + 1 e2x e2x − 21 e4x + 21 e2x ⎥
⎣ 2 2 ⎦.
− 2 e + 21 e2x
1 4x
−e4x + e2x 2e + 2e
1 4x 1 2x
⎡ ⎤
( 23 γ1 + γ2 + 21 γ3 )e4x + (− 21 γ1 − γ2 − 21 γ3 )e2x
⎢ ⎥
Y =⎢
⎣ (− 2 γ1 − 2 γ3 )e + ( 2 γ1 + γ2 + 2 γ3 )e
1 1 4x 1 1 2x ⎥.
⎦
(− 2 γ1 − γ2 + 2 γ3 )e + ( 2 γ1 + γ2 + 2 γ3 )e
1 1 4x 1 1 2x
⎡ ⎤
e−2x − xe−2x xe−2x −xe−2x
13a. eAx = ⎣e −2x − xe−2x − e4x xe −2x + e4x −xe−2x ⎦ .
e−2x − e4x −e−2x + e4x e−2x
⎡ ⎤
γ1 e−2x + (−γ1 + γ2 − γ3 )xe−2x
Y = ⎣γ1 e−2x + (−γ1 + γ2 − γ3 )xe−2x + (−γ1 + γ2 )e4x ⎦ .
(γ1 − γ2 + γ3 )e−2x + (−γ1 + γ2 )e4x
⎡ ⎤
3 − 2e3x 9x 2 + 6x − 2e3x
15a. eAx = ⎣ 2 − 2e3x 1 + 6x 2 + 4x − 2e3x ⎦ .
−3 + 3e3x −9x −2 − 6x + 3e3x
⎡ ⎤
(3γ1 + 2γ3 ) + (9γ2 + 6γ3 )x + (−2γ1 − 2γ3 )e3x
Y = ⎣(2γ1 + γ2 + 2γ3 ) + (6γ2 + 4γ3 )x + (−2γ1 − 2γ3 )e3x ⎦ .
(−3γ1 − 2γ3 ) + (−9γ2 − 6γ3 )x + (3γ1 + 3γ3 )e3x
⎡ ⎤
ex − 2xex xex −xex
17a. eAx = ⎣ −10xex ex + 5xex −5xex ⎦ .
−6xex 3xex e − 3xex
x
84 APPENDIX B. ANSWERS TO ODD-NUMBERED EXERCISES
⎡ ⎤
γ1 ex + (−2γ1 + γ2 − γ3 )xex
Y = ⎣γ2 ex + (−10γ1 + 5γ2 − 5γ3 )xex ⎦ .
γ3 ex + (−6γ1 + 3γ2 − 3γ3 )xex
⎡ ⎤
1 − 4x − 23 x 2 x + 21 x 2 2x + 21 x 2
19a. eAx = ⎣ −5x − 3x 2 1 + x + x2 3x + x 2 ⎦ .
−7x − 23 x 2 2x + 21 x 2 1 + 3x + 21 x 2
⎡ ⎤
γ1 + (−4γ1 + γ2 + 2γ3 )x + (− 23 γ1 + 21 γ2 + 21 γ3 )x 2
Y = ⎣ γ2 + (−5γ1 + γ2 + 3γ3 )x + (−3γ1 + γ2 + γ3 )x 2 ⎦ .
γ3 + (−7γ1 + 2γ2 + 3γ3 )x + (− 23 γ1 + 21 γ2 + 21 γ3 )x 2
e4x (cos(3x) − 13 sin(3x)) 5 4x
e sin(3x)
21a. eAx = 3 .
− 23 e4x sin(3x) e4x (cos(3x) + 13 sin(3x))
γ1 e4x cos(3x) + (− 13 γ1 + 53 γ2 )e4x sin(3x)
Y = .
γ2 e4x cos(3x) + (− 23 γ1 + 13 γ2 )e4x sin(3x)
e7x (cos(3x) − 2
3 sin(3x)) 13 7x
3 e sin(3x)
23a. eAx = .
− 13 e7x sin(3x) e7x (cos(3x) + 2
3 sin(3x))
γ1 e7x cos(3x) + (− 23 γ1 + 13 7x
3 γ2 )e sin(3x)
Y = .
γ2 e7x cos(3x) + (− 13 γ1 + 23 γ2 )e7x sin(3x)
85
Index