Jordan Canonical Form

Jordan Canonical Form:
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.
Jordan Canonical Form: Application to Differential Equations

Steven H. Weintraub
www.morganclaypool.com
ISBN: 9781598298048 paperback

ISBN: 9781598298055 ebook
DOI 10.2200/S00146ED1V01Y200808MAS002
A Publication in the Morgan & Claypool Publishers series

SYNTHESIS LECTURES ON MATHEMATICS AND STATISTICS
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
Steven H. Weintraub
Lehigh University
SYNTHESIS LECTURES ON MATHEMATICS AND STATISTICS #2
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
1 Jordan Canonical Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 The Diagonalizable Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Solving Systems of Linear Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.1 Homogeneous Systems with Constant Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . .25
2.2 Homogeneous Systems with Constant Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . .40
2.3 Inhomogeneous Systems with Constant Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.4 The Matrix Exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
A Background Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
A.1 Bases, Coordinates, and Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
A.2 Properties of the Complex Exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
B Answers to Odd-Numbered Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
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.
As is customary in textbooks, we provide the answers to the odd-numbered exercises here.

Instructors may contact me at shw2@lehigh.edu and I will supply the answers to all of the exercises.
Steven H. Weintraub
Lehigh University
Bethlehem, PA USA
July 2008
viii PREFACE
1
CHAPTER 1
Jordan Canonical Form

1.1 THE DIAGONALIZABLE CASE
Although, for simplicity, most of our examples will be over the real numbers (and indeed over the
rational numbers), we will consider that all of our vectors and matrices are defined over the complex
numbers C. It is only with this assumption that the theory of Jordan Canonical Form ( JCF) works
completely. See Remark 1.9 for the key reason why.
Definition 1.1. If v = 0 is a vector such that, for some λ,
Av = λv ,
then v is an eigenvector of A associated to the eigenvalue λ.

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.
We note that the definition of an eigenvalue/eigenvector can be expressed in an alternate form.

Here I denotes the identity matrix:
Av = λv
Av = λI v
(A − λI )v = 0 .
For an eigenvalue λ of A, we let Eλ denote the eigenspace of λ,
Eλ = {v | Av = λv} = {v | (A − λI )v = 0} = Ker(A − λI ) .
(The kernel Ker(A − λI ) is also known as the nullspace NS(A − λI ).)

We also note that this alternate formulation helps us find eigenvalues and eigenvectors. For if
(A − λI )v = 0 for a nonzero vector v, the matrix A − λI must be singular, and hence its determinant
must be 0. This leads us to the following definition.
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
We now introduce two important quantities associated to an eigenvalue of a matrix A.
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.
Lemma 1.7. Let a be an eigenvalue of a matrix A. Then
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.
Proof. In this case, applying Lemma 1.7, we have
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.
Lemma 1.10. Let A be an n-by-n matrix. The following are equivalent:

(1) For each eigenvalue a of A, geom-mult(a) = alg-mult(a).
(2) The sum of the geometric 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.
Proof. (1) Suppose we have a linear combination 0 = c1 v1 + c2 v2 + . . . + cm vm . We need to show

that ci = 0 for each i. To do this, we begin with an observation: If v is an eigenvector of A associated
to the eigenvalue a, and b is any scalar, then (A − bI )v = Av − bv = av − bv = (a − b)v. (Note
that this answer is 0 if a = b and nonzero if a = b.)
We now go to work, multiplying our original relation by (A − am I ). Of course, (A − am I )0 =
0, so:
0 = (A − am I )(c1 v1 + c2 v2 + . . . + cm−2 vm−2 + cm−1 vm−1 + cm vm )

= c1 (A − am I )v1 + c2 (A − am I )v2 + . . .
+ cm−2 (A − am I )vm−2 + cm−1 (A − am I )vm−1 + cm (A − am I )vm
= c1 (a1 − am )v1 + c2 (a2 − am )v2 + . . .
+ cm−2 (am−2 − am )vm−2 + cm−1 (am−1 − am )vm−1 .
We now multiply this relation by (A − am−1 I ). Again, (A − am−1 I )0 = 0, so:
0 = (A − am−1 I )(c1 (a1 − am )v1 + c2 (a2 − am )v2 + . . .

+ cm−2 (am−2 − am )vm−2 + cm−1 (am−1 − am )vm−1 )
= c1 (a1 − am )(A − am−1 I )v1 + c2 (a2 − am )(A − am−1 I )v2 + . . .
+ cm−2 (am−2 − am )(A − am−1 I )vm−2 + cm−1 (am−1 − am )(A − am−1 I )vm−1
= c1 (a1 − am )(a1 − am−1 )v1 + c2 (a2 − am )(a2 − am−1 )v2 + . . .
+ cm−2 (am−2 − am )(am−2 − am−1 )vm−2 .
Proceed in this way, until at the last step we multiply by (A − a2 I ). We then obtain:
0 = c1 (a1 − a2 ) · · · (a1 − am−1 )(a1 − am )v1 .
But v1 = 0, as by definition an eigenvector is nonzero. Also, the product (a1 − a2 ) · · · (a1 −

am−1 )(a1 − am ) is a product of nonzero numbers and is hence nonzero. Thus, we must have c1 = 0.
Proceeding in the same way, multiplying our original relation by (A − am I ), (A − am−1 I ),
(A − a3 I ), and finally by (A − a1 I ), we obtain c2 = 0, and, proceeding in this vein, we obtain ci = 0
for all i, and so the set {v1 , v2 , . . . , vm } is linearly independent.
(2) To avoid complicated notation, we will simply prove this when m = 2 (which illustrates the
general case). Thus, let m = 2, let S1 = {v1,1 , . . . , v1,i1 } be a linearly independent set of eigenvectors
associated to the eigenvalue a1 of A, and let S2 = {v2,1 , . . . , v2,i2 } be a linearly independent set
of eigenvectors associated to the eigenvalue a2 of A. Then S = {v1,1 , . . . , v1,i1 , v2,1 , . . . , v2,i2 }.
We want to show that S is a linearly independent set. Suppose we have a linear combination
0 = c1,1 v1,1 + . . . + c1,i1 v1,i1 + c2,1 v2,1 + . . . + c2,i2 v2,i2 . Then:
0 = c1,1 v1,1 + . . . + c1,i1 v1,i1 + c2,1 v2,1 + . . . + c2,i2 v2,i2

= (c1,1 v1,1 + . . . + c1,i1 v1,i1 ) + (c2,1 v2,1 + . . . + c2,i2 v2,i2 )
= v1 + v 2
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),
0 = (A − a2 I )0 = (A − a2 I )(v1 + v2 ) = (A − a2 I )v1 + (A − a2 I )v2

= (a1 − a2 )v1 + 0 = (a1 − a2 )v1
so 0 = v1 ; similarly, 0 = v2 . But 0 = v1 = c1,1 v1,1 + . . . + c1,i1 v1,i1 implies c1,1 = . . . c1,i1 = 0,

as, by hypothesis, {v1,1 , . . . , v1,i1 } is a linearly independent set; similarly, 0 = v2 implies c2,1 =
. . . = c2,i2 = 0. Thus, c1,1 = . . . = c1,i1 = c2,1 = . . . = c2,i2 = 0 and S is linearly independent, as
claimed. 2
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.
Here is the main result of this section.
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.
First let us suppose that for each eigenvalue a of A, geom-mult(a) = alg-mult(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
and we see that ji , the i th column of J , is the vector

⎡ ⎤
0
⎢ .. ⎥
⎢.⎥
⎢ ⎥
⎢0⎥
ji = ⎢ ⎥
⎢ai ⎥ ,
⎢ ⎥
⎢0⎥
⎣ ⎦
..
.
with ai in the i th position, and 0 elsewhere.

We have

P = v1 v2 . . . vn ,

a matrix whose columns are eigenvectors forming bases for the associated eigenspaces. By hypothesis,
geom-mult(a) = alg-mult(a) for each eigenvector a of A, so there are as many columns of P that are
eigenvectors for the eigenvalue a as there are diagonal entries of J that are equal to a. Furthermore,
by Lemma 1.10, the matrix P indeed has n columns.
We first show by direct computation that AP = P J . Now

AP = A v1 v2 . . . vn
so the i th column of AP is Avi . But

Avi = ai vi
as vi is an eigenvector of A with associated eigenvalue ai .
On the other hand,

P J = v1 v2 . . . vn J
and the i th column of P J is P ji ,

P ji = v1 v2 . . . vn ji .

Remembering what the vector ji is, and multiplying, we see that
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 ,
completing the proof of this half of the Theorem.

Now let us suppose that A is diagonalizable, A = P J P −1 . Then AP = P J . We use the same
notation for P and J as in the first half of the proof. Then, as in the first half of the proof, we
compute AP and P J column-by-column, and we see that the i th column of AP is Avi and that
the i th column of P J is ai vi , for each i. Hence, Avi = ai vi for each i, and so vi is an eigenvector
of A with associated eigenvalue ai .
For each eigenvalue a of A, there are as many columns of P that are eigenvectors for a as
there are diagonal entries of J that are equal to a, and these vectors form a basis for the eigenspace
associated of the eigenvalue a, so we see that for each eigenvalue a of A, geom-mult(a) = alg-mult(a),
completing the proof. 2
For a general matrix A, the condition in Theorem 1.14 may or may not be satisfied, i.e.,
some but not all matrices are diagonalizable. But there is one important case when this condition is
automatic.
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.
1.2 THE GENERAL CASE

Let us begin this section by describing what a matrix in JCF looks like. A matrix in JCF is composed
of “Jordan blocks,” so we first see what a single Jordan block looks like.
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
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
with each Ji a Jordan block.
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.
Definition 2.4. If v = 0 is a vector such that, for some λ,
(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.
Let us note that if v is a generalized eigenvector of index 1, then
(A − λI )(v) = 0
(A)v = (λI )v
Av = λv
and so v is an (ordinary) eigenvector.

Recall that, for an eigenvalue λ of A, Eλ denotes the eigenspace of λ,
Eλ = {v | Av = λv} = {v | (A − λI )v = 0} .
We let E˜λ denote the generalized eigenspace of λ,
E˜λ = {v | (A − λI )k (v) = 0 for some k} .
It is easy to check that E˜λ is a subspace.

Since every eigenvector is a generalized eigenvector, we see that
Eλ ⊆ E˜λ .
The following result (which we shall not prove) is an important fact about generalized
eigenspaces.
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 ).
Let us next consider a generalized eigenvector vk of index k associated to an eigenvalue λ, and

set
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
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:
Definition 2.7. If {v1 , . . . , vk } is a set of generalized eigenvectors associated to the eigenvalue λ of

A, such that vk is a generalized eigenvector of index k and also
vk−1 =(A − λI )vk , vk−2 = (A − λI )vk−1 , vk−3 = (A − λI )vk−2 ,

· · · , v2 = (A − λI )v3 , v1 = (A − λI )v2 ,
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
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.
Proof. Suppose we have a linear combination
c1 v1 + c2 v2 + · · · + ck−1 vk−1 + ck vk = 0 .
We must show each ci = 0.

By the definition of a chain, vk−i = (A − λI )i vk for each i, so we may write this equation as
c1 (A − λI )k−1 vk + c2 (A − λI )k−2 vk + · · · + ck−1 (A − λI )vk + ck vk = 0 .

Now let us multiply this equation on the left by (A − λI )k−1 . Then we obtain the equation
c1 (A − λI )2k−2 vk + c2 (A − λI )2k−3 vk + · · · + ck−1 (A − λI )k vk + ck (A − λI )k−1 vk = 0 .
Now (A − λI )k−1 vk = v1 = 0. However, (A − λI )k vk = 0, and then also (A − λI )k+1 vk =

(A − λI )(A − λI )k vk = (A − λI )(0) = 0, and then similarly (A − λI )k+2 vk = 0, . . . , (A −
λI )2k−2 vk = 0, so every term except the last one is zero and this equation becomes
ck v1 = 0 .
Since v1 = 0, this shows ck = 0, so our linear combination is
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 ,
so {v1 , . . . , vk } is linearly independent. 2
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

is a matrix whose columns are generalized eigenvectors forming a chain.

Proof. We give a proof by direct computation here. (Note the similarity of this proof to the proof
of Theorem 1.14.) For a more conceptual proof, see Theorem 1.11 in Appendix A.
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

and we see that ji , the i th column of J , is the vector

⎡ ⎤
0
⎢ .. ⎥
⎢.⎥
⎢ ⎥
⎢1⎥
ji = ⎢
⎢a ⎥
⎥
⎢ ⎥
⎢0 ⎥
⎣ ⎦
..
.
with 1 in the (i − 1)st position, a in the i th position, and 0 elsewhere.

We show that AP = P J by showing that their corresponding columns are equal.
Now

AP = A v1 v2 . . . vk
so the i th column of AP is Avi . But
Avi = (A − aI + aI )vi
= (A − aI )vi + aI vi
= vi−1 + avi for i > 1, = avi for i = 1 .
On the other hand,

PJ = v1 v2 . . . vk J

and the i th column of P J is P ji ,

P ji = v1 v2 . . . vk ji .

Remembering what the vector ji is, and multiplying, we see that
P ji = vi−1 + avi for i > 1, = avi for i = 1
as well.
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:
(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
(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
(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
(iii) A has a single eigenvalue a of algebraic multiplicity 3.
(a) A has three linearly independent eigenvectors, u, v, and w. Then A = P J P −1 with

⎡ ⎤
a 0 0

J = ⎣0 a 0⎦ and P = u v w .
0 0 a
(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
(c) A has a single chain {u1 , u2 , u3 } of generalized eigenvectors. Then A = P J P −1 with

⎡ ⎤
a 1 0

⎣
J = 0 a 1 ⎦ and P = u1 u2 u3 .
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.
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 ).
Definition 2.18. Let λ be an eigenvalue of A. Then for any positive integer i,
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.
(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.
Now we shall do several examples.

⎡ ⎤
2 −3 −3
Example 2.23. A = ⎣ 2 −2 −2⎦ .
−2 1 1
A has characteristic polynomial det (λI − A) = (λ + 1)(λ)(λ − 2). Thus, A has eigenvalues
−1, 0, and 2, each of multiplicity one, and so we are in the situation
⎧⎡ ⎤ of⎫Theorem 2.15 (i). Computation
⎨ 1 ⎬
shows that the eigenspace E−1 = Ker(A − (−I )) has basis ⎣0⎦ , the eigenspace E0 = Ker(A)
⎩ ⎭
1
⎧⎡ ⎤⎫ ⎧⎡ ⎤⎫
⎨ 0 ⎬ ⎨ −1 ⎬
has basis ⎣ 1 ⎦ , and the eigenspace E2 = Ker(A − 2I ) has basis ⎣−1⎦ . Hence, we see
⎩ ⎭ ⎩ ⎭
−1 1
that
⎡ ⎤ ⎡ ⎤⎡ ⎤⎡ ⎤−1
2 −3 −3 1 0 −1 −1 0 0 1 0 −1
⎣ 2 −2 −2⎦ = ⎣0 −1 −1⎦ ⎣ 0 0 0⎦ ⎣0 −1 −1⎦ .
−2 1 1 1 1 1 0 0 2 1 1 1
⎡ ⎤
3 1 1
Example 2.24. A = ⎣2 4 2⎦ .
1 1 3
A has characteristic polynomial det (λI − A) = (λ − 2)2 (λ − 6). Thus, A

has an eigenvalue
2 of multiplicity 2 and an eigenvalue
⎧⎡ ⎤ 6 of multiplicity 1. Computation shows that the eigenspace
⎡ ⎤⎫
⎨ −1 −1 ⎬
E2 = Ker(A − 2I ) has basis ⎣ 1 ⎦ , ⎣ 0 ⎦ , so dim E2 = 2 and we are in the situation of
⎩ ⎭
0 1
Corollary
⎧ 2.21 (a). Further computation shows that the eigenspace E6 = Ker(A − 6I ) has basis
⎡ ⎤⎫
⎨ 1 ⎬
⎣2⎦ . Hence, we see that
⎩ ⎭
1
⎡ ⎤ ⎡ ⎤⎡ ⎤⎡ ⎤−1
3 1 1 −1 −1 1 2 0 0 −1 −1 1
⎣2 4 2 ⎦ = ⎣ 1 0 2⎦ ⎣0 2 0⎦ ⎣ 1 0 2⎦ .
1 1 3 0 1 1 0 0 6 0 1 1
⎡ ⎤
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
⎡ ⎤
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 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.
EXERCISES FOR CHAPTER 1

For each matrix A, write A = P J P −1 with P an invertible matrix and J a matrix in JCF.

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

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
⎡ ⎤
−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
Solving Systems of Linear

2.1 HOMOGENEOUS SYSTEMS WITH CONSTANT
COEFFICIENTS
We will now see how to use Jordan Canonical Form ( JCF) to solve systems Y = AY . We begin by
describing the strategy we will follow throughout this section.
Consider the matrix system

Y = AY .
Step 1. Write A = P J P −1 with J in JCF, so the system becomes
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 ) .
(Note that, since P −1 is a constant matrix, we have that (P −1 Y ) = P −1 Y .)
Step 2. Set Z = P −1 Y , so this system becomes
Z = J Z
and solve this system for Z.
Step 3. Since Z = P −1 Y , we have that
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.
Theorem 1.1. Let J be a k-by-k diagonal matrix,

⎡ ⎤
a1
⎢ a2 0 ⎥
⎢ ⎥
⎢ a3 ⎥
⎢ ⎥
J =⎢ .. ⎥.
⎢ . ⎥
⎢ ⎥
⎣ 0 ak−1 ⎦
ak
Then the system Z = J Z has the solution

⎡ ⎤
ea1 x
⎢ ea2 x 0 ⎥
⎢ ⎥
⎢ ea3 x ⎥
⎢ ⎥
Z=⎢ .. ⎥ C = MZ C
⎢ . ⎥
⎢ ⎥
⎣ 0 eak−1 x ⎦
eak x
⎡ ⎤
c1
⎢ c2 ⎥
⎢ ⎥
where C = ⎢ . ⎥ is a vector of arbitrary constants c1 , c2 , . . . , ck .
⎣ .. ⎦
ck
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
which is exactly the above product MZ C. 2
Example 1.2. Consider the system

5 −7
Y = AY where A= .
2 −4
We saw in Example 1.16 in Chapter 1 that A = P J P −1 with

7 1 3 0
P = and J = .
2 1 0 −2
Then Z = J Z has solution

e3x 0 c1 c1 e3x
Z= = M C =
0 e−2x c2 Z
c2 e−2x
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

⎡ ⎤
2 −3 −3
Y = AY where A = ⎣ 2 −2 −2⎦ .
−2 1 1

⎡ ⎤ ⎡ ⎤
1 0 −1 −1 0 0
P = ⎣0 −1 −1⎦ and J = ⎣ 0 0 0⎦ .
1 1 1 0 0 2

⎡ ⎤⎡ ⎤
e−x 0 0 c1
Z= ⎣ 0 1 0 ⎦ ⎣c2 ⎦ = MZ C
0 0 e2x c3
⎡ ⎤ ⎡ −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.
Theorem 1.4. Let J be a k-by-k Jordan block with eigenvalue a,

⎡ ⎤
a 1
⎢ a 1 0 ⎥
⎢ ⎥
⎢ a 1 ⎥
⎢ ⎥
J =⎢ .. .. ⎥.
⎢ . . ⎥
⎢ ⎥
⎣ 0 a 1⎦
a
Then the system Z = J Z has the solution
⎡ ⎤
1 x x 2 /2! x 3 /3! · · · x k−1 /(k − 1)!
⎢ 1 x x 2 /2! · · · x k−2 /(k − 2)!⎥
⎢ ⎥
⎢ 1 x · · · x k−3 /(k − 3)!⎥
ax ⎢ ⎥
Z=e ⎢ .. .. ⎥ C = MZ C
⎢ . . ⎥
⎢ ⎥
⎣ x ⎦
1
⎡ ⎤
c1
⎢c2 ⎥
⎢ ⎥
where C = ⎢ . ⎥ is a vector of arbitrary constants c1 , c2 , . . . , ck .
⎣ .. ⎦
ck
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 ]
which is nothing other than the differential equation
z1 = az1 .
This differential equation has solution
z1 = c1 eax ,
which we can certainly write as

[z1 ] = eaz [1][c1 ] .

z1 a 1 z1
= ,
z2 0 a z2
which is nothing other than the pair of differential equations
z1 = az1 + z2
z2 = az2 .
We recognize the second equation as having the solution
z2 = c2 eax
and we substitute this into the first equation to get
z1 = az1 + c2 eax .
To solve this, we rewrite this as
z1 − az1 = c2 eax
and recognize that this differential equation has integrating factor e−ax . Multiplying by this factor,
we find
e−ax (z1 − az1 ) = c2

(e−ax z1 ) = c2
e−ax z1 = c2 dx = c1 + c2 x
so
z1 = eax (c1 + c2 x) .
Thus, our solution is
z1 = eax (c1 + c2 x)
z2 = eax c2 ,
which we see we can rewrite as

z1 ax 1 x c1
=e .
z2 0 1 c2

⎡ ⎤ ⎡ ⎤⎡ ⎤
z1 a 1 0 z1
⎣z ⎦ = ⎣0 a 1⎦ ⎣z2 ⎦ ,
2
z3 0 0 a z3
which is nothing other than the triple of differential equations
z1 = az1 + z2
z2 = az2 + z3
z3 = az3 .
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 .
Substituting the value of z2 into the equation for z1 , we obtain
z1 = az1 + eax (c2 + c3 x) .
To solve this, we rewrite this as
z1 − az1 = eax (c2 + c3 x)
and recognize that this differential equation has integrating factor e−ax . Multiplying by this factor,
we find
e−ax (z1 − az1 ) = c2 + c3 x

(e−ax z1 ) = c2 + c3 x
e−ax z1 = (c2 + c3 x) dx = c1 + c2 x + c3 (x 2 /2)
so
z1 = eax (c1 + c2 x + c3 (x 2 /2)) .
Thus, our solution is
z1 = eax (c1 + c2 x + c3 (x 2 /2))

z2 = eax (c2 + c3 x)
z3 = eax c3 ,
which we see we can rewrite as

⎡ ⎤ ⎡ ⎤⎡ ⎤
z1 1 x x 2 /2 c1
⎣z2 ⎦ = eax ⎣0 1 x ⎦ ⎣c2 ⎦ .
z3 0 0 1 c3
2
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.
We now illustrate this (confining our illustrations to the case that A is not diagonalizable, as
we have already illustrated the diagonalizable case).

0 1
Y = AY where A= .
−4 4

−2 1 2 1
P = and J = .
−4 0 0 2

2x 2x
2x 1 x c1 e xe2x c1 c1 e + c2 xe2x
Z= e = = MZ C =
0 1 c2 0 e2x c2 c2 e2x

−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

⎡ ⎤
2 1 1
Y = AY where A = ⎣ 2 1 −2⎦ .
−1 0 −2

⎡ ⎤ ⎡ ⎤
1 0 −5 −1 1 0
P = ⎣−2 0 −6⎦ and J = ⎣ 0 −1 0⎦ .
−1 1 1 0 0 3
⎡ ⎤⎡ ⎤
e−x xe−x 0 c1
Z= ⎣ 0 e−x 0 ⎦ ⎣c2 ⎦ = MZ C
0 0 e3x c3
⎡ ⎤ ⎡ −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

⎡ ⎤
2 1 1
Y = AY where A = ⎣−2 −1 −2⎦ .
1 1 2

⎡ ⎤ ⎡ ⎤
1 1 1 1 1 0
P = ⎣−2 0 0⎦ and J = ⎣0 1 0⎦ .
1 0 1 0 0 1

⎡ ⎤⎡ ⎤
ex xex 0 c1
Z = ⎣0 ex 0 ⎦ ⎣c2 ⎦ = MZ C
0 0 ex c3
⎡ ⎤⎡ x ⎤⎡ ⎤
1 1 1 e xex 0 c1
⎣
Y = −2 0 0 ⎦ ⎣ 0 ex 0 ⎦ ⎣c2 ⎦
1 0 1 0 0 e x c3
⎡ ⎤⎡ ⎤
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

⎡ ⎤
5 0 1
Y = AY where A = ⎣ 1 1 0⎦ .
−7 1 0

⎡ ⎤ ⎡ ⎤
2 3 1 2 1 0
P =⎣ 2 1 0⎦ and J = ⎣0 2 1 ⎦ .
−6 −7 0 0 0 2

⎡ ⎤⎡ ⎤
e2x xe2x (x 2 /2)e2x c1
Z= ⎣ 0 e2x xe2x ⎦ ⎣c2 ⎦ = MZ C
0 0 e2x c3
⎡ ⎤ ⎡ 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
We conclude this section by showing how to solve initial value problems. This is just one more
step, given what we have already done.
Example 1.10. Consider the initial value problem

0 1 3
Y = AY where A = , and Y (0) = .
−4 4 −8
In Example 1.6, we saw that this system has the general solution

(−2c1 + c2 )e2x − 2c2 xe2x
Y = .
−4c1 e2x − 4c2 xe2x
Applying the initial condition (i.e., substituting x = 0 in this matrix), gives

3 −2c1 + c2
= Y (0) =
−8 −4c1
with solution
c1 2
= .
c2 7
Substituting these values in the above matrix gives
2x
3e − 14xe2x
Y = .
−8e2x − 28te2x

⎡ ⎤ ⎡ ⎤
2 1 1 8
Y = AY where A = ⎣ 2 1 −2⎦ , and Y (0) = ⎣32⎦ .
−1 0 2 5
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
Applying the initial condition (i.e., substituting x = 0 in this matrix) gives

⎡ ⎤ ⎡ ⎤
8 c1 − 5c3
⎣32⎦ = Y (0) = ⎣ −2c1 − 6c3 ⎦
5 −c1 + c2 + c3
with solution ⎡ ⎤ ⎡ ⎤
c1 −7
⎣c2 ⎦ = ⎣ 1 ⎦ .
c3 −3
Substituting these values in the above matrix gives
⎡ ⎤
−7e−x + xe−x + 15e3x
Y = ⎣14e−x − 2xe−x + 18e3x ⎦ .
8e−x − xe−x − 3e3x
Remark 1.12. There is a variant on our method of solving systems or initial value problems.
We have written our solution of Z = J Z as Z = MZ C. Let us be more explicit here and

write this solution as
Z(x) = MZ (x)C .
This notation reminds us that Z(x) is a vector of functions, MZ (x) is a matrix of functions, and C is
a vector of constants. The key observation is that MZ (0) = I , the identity matrix. Thus, if we wish
to solve the initial value problem
Z = J Z, Z(0) = Z0 ,
we find that, in general,

Z(x) = MZ (x)C
and, in particular,
Z0 = Z(0) = MZ (0)C = I C = C ,
so the solution to this initial value problem is
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 .
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.
Let us now apply this variant method.

0 1 a1
Y = AY where A = , and Y (0) =
.
−4 4 a2

−1 −2 1 2 1
As we have seen in Example 1.6, A = P J P with P = and J = .Then MZ (x) =
−4 0 0 2
2x
e xe2x
and
0 e2x
−1
−2 1 e2x xe2x −2 1
MY (x) = P MZ (x)P −1 =
−4 0 0 e2x −4 0
2x
e − 2xe 2x xe 2x
=
−4xe2x e2x + 2xe2x
so
2x
a e − 2xe2x xe2x a1
Y (x) = MY (x) 1 =
a2 −4xe2x e2x + 2xe2x a2

a1 e2x + (−2a1 + a2 )xe2x
= .
a2 e2x + (−4a1 + 2a2 )xe2x
2x
3 3e − 14xe2x
In particular, if Y (0) = , then Y (x) = , recovering the result of Exam-
−8 −8e2x − 28xe2x
2x
2 2e + xe2x −4
ple 1.10. But also, if Y (0) = , then Y (x) = , and if Y (0) = , then Y (x) =
5 5e + 2te
2x 2x 15

−4e2x + 23xe2x
, etc.
15e2x + 46xe2x
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 .
EXERCISES FOR SECTION 2.1

For each exercise, see the corresponding exercise in Chapter 1. In each exercise:
(a) Solve the system Y = AY .
(b) Solve the initial value problem Y = AY , Y (0) = Y0 .

75 56 1
1. A = and Y0 = .
−90 −67 −1

−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
⎡ ⎤ ⎡ ⎤
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
⎡ ⎤ ⎡ ⎤
−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
2.2 HOMOGENEOUS SYSTEMS WITH CONSTANT

COEFFICIENTS: COMPLEX ROOTS
In this section, we show how to solve a homogeneous system Y = AY where the characteristic
polynomial of A has complex roots. In principle, this is the same as the situation where the
characteristic polynomial of A has real roots, which we dealt with in Section 2.1, but in practice,
there is an extra step in the solution.
We will begin by doing an example, which will show us where the difficulty lies, and then we
will overcome that difficulty. But first, we need some background.
Definition 2.1. For a complex number z, the exponential ez is defined by
ez = 1 + z + z2 /2! + z3 /3! + . . . .
The complex exponential has the following properties.
Theorem 2.2. (1) (Euler) For any θ,
eiθ = cos(θ ) + i sin(θ ) .
(2) For any a,

d az
(e ) = aeaz .
dz
(3) For any z1 and z2 ,
ez1 +z2 = ez1 ez2 .
(4) If z = s + it, then
ez = es (cos(t) + i sin(t)) .
(5) For any z,
ez = ez .
Proof. For the proof, see Theorem 2.2 in Appendix A. 2

The following lemma will save us some computations.
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
Now for our example.

2 −17
Y = AY where A= .
1 4
A has characteristic polynomial λ2 − 6λ + 25 with roots λ1 = 3 + 4i and λ2 = λ1 = 3 − 4i, each

of multiplicity 1. Thus, λ1 and λ2 are the eigenvalues of A, and we compute that the eigenspace

−1 + 4i
E3+4i = Ker(A − (3 + 4i)I ) has basis v1 = , and hence, by Lemma 2.3, that the
1

−1 − 4i
eigenspace E3−4i = Ker(A − (3 − 4i)I ) has basis v2 = v1 = . Hence, just as before,
1

−1 −1 + 4i −1 − 4i 3 + 4i 0
A = PJP with P = and J = .
1 1 0 3 − 4i
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
where f1 is an arbitrary complex constant.

This solution is correct but unacceptable. We want to solve the system Y = AY , where A has
real coefficients, and we have a solution which is indeed a real vector, but this vector is expressed in
terms of complex numbers and functions. We need to obtain a solution that is expressed totally in
terms of real numbers and functions. In order to do this, we need an extra step.
In order not to interrupt the flow of exposition, we simply state here what we need to do, and
we justify this after the conclusion of the example.
We therefore do the following: We simply replace the matrix P MZ by the matrix whose

−1 + 4i
first column is the real part Re(e v1 ) = Re e
λ1 x (3+4i)x , and whose second column is
1

−1 + 4i
the imaginary part Im(eλ1 x v1 ) = Im e(3+4i)x , and the vector F by the vector C of
1
arbitrary real constants. We compute

(3+4i)x −1 + 4i −1 + 4i
e = e (cos(4x) + i sin(4x))
3x
1 1

− cos(4x) − 4 sin(4x) 3x 4 cos(4x) − sin(4x)
=e 3x
+ ie
cos(4x) sin(4x)
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)
Now we justify the step we have done.
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
where c1 and c2 are arbitrary real constants.
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
Now Y = AY has general solution Y = P MZ F = P MZ (RR −1 )F = (P MZ R)(R −1 F ) for

any invertible matrix R. We now (cleverly) choose

1/2 1/(2i)
R= .
1/2 −1/(2i)
With this choice of R,

P MZ R = Re(eλ1 x v1 ) Im(eλ1 x v1 ) .
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
yielding a general solution as claimed. 2

We now solve Y = AY where A is a real 3-by-3 matrix with a pair of complex eigenvalues
and a third, real eigenvalue. As you will see, we use the idea of Lemma 2.5 to simply replace the
“relevant” columns of P MZ in order to obtain our final solution.

⎡ ⎤
15 −16 8
Y = AY where A = ⎣10 −10 5⎦ .
0 1 2
A has characteristic polynomial (λ2 − 2λ + 5)(λ − 5) with roots λ1 = 1 + 2i, λ2 = λ1 = 1 − 2i,

and λ3 = 5, each of multiplicity 1. Thus, λ1 , λ2 , and λ3 are the
⎧ eigenvalues of A,⎫and we compute
⎡ ⎤
⎨ −2 + 2i ⎬
that the eigenspace E1+2i = Ker(A − (1 + 2i)I ) has basis v1 = ⎣−1 + 2i ⎦ , and hence, by
⎩ ⎭
1
⎧ ⎡ ⎤⎫
⎨ −2 − 2i ⎬
Lemma 2.3, that the eigenspace E1−2i = Ker(A − (1 − 2i)I ) has basis v2 = v1 = ⎣−1 − 2i ⎦ .
⎩ ⎭
1
⎧ ⎡ ⎤⎫
⎨ 4 ⎬
We further compute that the eigenspace E5 = Ker(A − 5I ) has basis v3 = 3⎦ . Hence, just as
⎣
⎩ ⎭
1
before,
⎡ ⎤ ⎡ ⎤
−2 + 2i −2 − 2i 4 1 + 2i 0 0
A = P J P −1 with P = ⎣−1 + 2i −1 − 2i 3⎦ and J = ⎣ 0 1 − 2i 0⎦ .
1 1 1 0 0 5

⎡ ⎤⎡ ⎤ ⎡ ⎤
e(1+2i)x 0 0 f1 f1 e(1+2i)x
⎢ ⎥
Z=⎣ 0 e(1−2i)x 0 ⎦ ⎣f1 ⎦ = MZ F = ⎣f1 e(1+2i)x ⎦
0 0 e5x c3 c3 e5x
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
= ⎣ (−c1 + 2c2 )ex cos(2x) + (−2c1 − c2 )ex sin(2x) + 3c3 e5x ⎦ .

c1 ex cos(2x) + c2 ex sin(2x) + c3 e5x

In Exercises 1–4:
(a) Solve the system Y = AY .
(b) Solve the initial value problem Y = AY , Y (0) = Y0 .

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
2.3 INHOMOGENEOUS SYSTEMS WITH CONSTANT

COEFFICIENTS
In this section, we show how to solve an inhomogeneous system Y = AY + G(x) where G(x)
is a vector of functions. (We will often abbreviate G(x) by G). We use a method that is a direct
generalization of the method we used for solving a homogeneous system in Section 2.1.
Consider the matrix system
Y = AY + G .
2.3. INHOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 47
Step 1. Write A = P J P −1 with J in JCF, so the system becomes
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 .
(Note that, since P −1 is a constant matrix, we have that (P −1 Y ) = P −1 Y .)
Step 2. Set Z = P −1 Y and H = P −1 G, so this system becomes
Z = J Z + H
and solve this system for Z.
Step 3. Since Z = P −1 Y , we have that
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 .
Proof. First we check that Y = Yh + Yi is a solution of Y = AY + G(x). We simply compute
Y = (Yh + Yi ) = Yh + Yi = (AYh ) + (AYi + G)

= A(Yh + Yi ) + G = AY + G
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
Yh = (Y − Yi ) = Y − Yi = (AY + G) − (AYi + G)

= A(Y − Yi ) = AYh
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.)
Theorem 3.2. The function

Yi = NY NY−1 G dx
0
is a solution of the system Y = AY + G.
Proof. We simply compute Yi . We have

Yi = NY NY−1 G dx
0
= NY NY−1 G dx + NY

NY−1 G dx
0
by the product rule
= NY NY−1 G dx + NY (NY−1 G)
0
by the definition of the antiderivative
= NY NY−1 G dx + G
0
= (ANY ) NY−1 G dx + G
0
as NY = ANY

= A NY NY−1 G dx + G
0
= AYi + G
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 .

5 −7 30ex
Y = AY + G where A= and G = .
2 −4 60e2x
We saw in Example 1.2 that

7 1 e3x 0
P = and MZ = ,
2 1 0 e−2x
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
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

3x
0 1 60e
−4 4 72e5x

−2 1 1 x
P = and MZ = e 2x
,
4 0 0 1

−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

⎡ ⎤ ⎡ x ⎤
2 −3 −3 e
Y = AY + G where ⎣ ⎦ ⎣
A = 2 −2 −2 and G = 12e3x ⎦ .
−2 1 1 20e4x
⎡ ⎤ ⎡ ⎤
1 0 −1 e−x 0 0
P = ⎣0 −1 −1⎦ and MZ = ⎣ 0 1 0 ⎦ ,
1 1 1 0 0 e2x
⎡ 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

2 −17 200
1 4 160ex

−1 + 4i −1 − 4i e(3+4i)x 0
P = and MZ = ,
1 1 0 e(3−4i)x

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
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.

7 5 −7 30ex
Y = AY + G, Y (0) = , where A= and G = .
17 2 −4 60e2x
We saw in Example 1.2 that the associated homogenous system has general solution

7c1 e3x + c2 e−2x
YH =
2c1 e3x + c2 e−2x
and in Example 3.3 that the original system has a particular solution

−25ex + 105e2x
Yi = .
−10ex + 45e2x
Thus, our original system has general solution

7c1 e3x + c2 e−2x − 25ex + 105e2x
Y = YH + Yi = .
2c1 e3x + c2 e−2x − 10ex + 45e2x
We apply the initial condition to obtain the linear system

7c1 + c2 + 80 7
Y (0) = =
2c1 + c2 + 35 17
2.4. THE MATRIX EXPONENTIAL 53
with solution c1 = −11, c2 = 4. Substituting, we find that our initial value problem has solution

−77e3x + 4e−2x − 25ex + 105e2x
Y = .
−22e3x + 4e−2x − 10ex + 45e2x

In each exercise, find a particular solution Yi of the system Y = AY + G(x), where A is the matrix
of the correspondingly numbered exercise for Section 2.1, and G(x) is as given.
8x
2e
1. G(x) = .
3e4x
−7x
2e
2. G(x) = .
6e−8x
4x
e
3. G(x) = .
4e5x
6x
e
4. G(x) = .
9e8x
10x
9e
5. G(x) = .
25e12x
−x
5e
6. G(x) = .
12e2x
⎡ ⎤
1
7. G(x) = ⎣3e2x ⎦.
5e4x
⎡ ⎤
8
8. G(x) = ⎣3e3x ⎦.
3e5x
2.4 THE MATRIX EXPONENTIAL

In this section, we will discuss the matrix exponential and its use in solving systems Y = AY .
Our first task is to ask what it means to take a matrix exponential. To answer this, we are
guided by ordinary exponentials. Recall that, for any complex number z, the exponential ez is given
by
ez = 1 + z + z2 /2! + z3 /3! + z4 /4! + . . . .
With this in mind, we define the matrix exponential as follows.
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.
(1) Let A be a square matrix. Then the general solution of
Y = AY
is given by
Y = eAx
where is a vector of arbitrary constants.
(2) The initial value problem

Y = AY, Y (0) = Y0
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!
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.
Lemma 4.3. If J is a diagonal matrix,

⎡ ⎤
d1
⎢ d2 0 ⎥
⎢ ⎥
J =⎢ .. ⎥
⎣ 0 . ⎦
dn
then eJ x is the diagonal matrix

⎡ ⎤
ed1 x
⎢ ed2 x 0 ⎥
⎢ ⎥
eJ x = ⎢ .. ⎥.
⎣ 0 . ⎦
edn x
Proof. Suppose, for simplicity, that J is 2-by-2,

d 0
J = 1 .
0 d2
3
d 2 0 d 0
Then you can easily compute that J2 = 1 2 , J = 1
3 , and similarly, J k =
0 d2 0 d2 3
k
d1 0
for any k.
0 d2 k
Then, as in the proof of Theorem 4.2,
1 2 2 1 1
eJ x = I + J x + J x + J 3x3 + J 4x4 + . . .
2! 3! 4!
2
1 0 d 0 1 d1 0 1 d1 3 0
= + 1 x+ x 2
+ x3 + . . .
0 1 0 d2 2! 0 d2 2 3! 0 d2 3

1 + d1 x + 2!1 (d1 x)2 + 3!1 (d1 x)3 + . . . 0
=
0 1 + d2 x + 2!1 (d2 x)2 + 3!1 (d2 x)3 + . . .
which we recognize as
dx
e 1 0
= .
0 ed2 x
2
Example 4.4. We wish to find the general solution of Y = J Y where

3 0
J = .
0 −2
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
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!
Lemma 4.5. Let S and T be two matrices and suppose
S = P T P −1
for some invertible matrix P . Then
S k = P T k P −1 for every k
and
eS = P eT P −1 .
Proof. We simply compute
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
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

7 1 3 0
P = and J = .
2 1 0 −2
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

⎡ ⎤ ⎡ ⎤
1 0 −1 −1 0 0
P = ⎣0 −1 −1⎦ and J = ⎣ 0 0 0⎦ .
1 1 1 0 0 2
Then
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
= ⎣−1 + e2x 2 − e2x 1 − e2x ⎦

1 − e2x e−x − 2 + e2x e−x − 1 + e2x
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

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
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.
Theorem 4.9. Let J be a k-by-k Jordan block with eigenvalue a,

⎡ ⎤
a 1
⎢ a 1 0 ⎥
⎢ ⎥
⎢ a 1 ⎥
⎢ ⎥
J =⎢ .. .. ⎥.
⎢ . . ⎥
⎢ ⎥
⎣ 0 a 1⎦
a
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
Proof. First suppose that J is a 2-by-2 Jordan block,

a 1
J = .
0 a
3 4
a2 2a a 3a 2 a 4a 3
Then J2 = 2 ,J =
3
3 ,J =
4 ,…
0 a 0 a 0 a4
so

1 0 a 1 1 a2 2a 2 1 a3 3a 2 3 1 a4 4a 3 4
e Jx
= + x+ x + x + x + ...
0 1 0 a 2! 0 a2 3! 0 a3 4! 0 a4

m11 m12
= ,
0 m22
and we see that

1 1 1 1
m11 = m22 = 1 + ax + (ax)2 + (ax)3 + (ax)4 + (ax)5 + . . .
2! 3! 4! 5!
= eax ,
and
1 2 3 1 1
m12 = x + ax 2 + a x + a3x 4 + a4x 5 + . . .
2! 3! 4!
1 1 1
= x(1 + ax + (ax)2 + (ax)3 + (ax)4 + . . .) = xeax
2! 3! 4!
and so we conclude that ax
e xeax ax 1 x
e =
Jx
=e .
0 eax 0 1
Next suppose that J is a 3-by-3 Jordan block,
⎡ ⎤
a 1 0
J = ⎣ 0 a 1⎦ .
0 0 1
⎡ 2 ⎤ ⎡ 3 ⎤ ⎡ 4 ⎤
a 2a 1 a 3a 2 3a a 4a 3 6a 2
Then J 2 = ⎣ 0 a 2 2a ⎦, J 3 = ⎣ 0 a 3 3a 2 ⎦, J 4 = ⎣ 0 a 4 4a 3 ⎦,
0 0 a2 0 0 a3 0 0 a4
⎡ 5 ⎤
a 5a 4 10a 3
J 5 = ⎣ 0 a5 5a 4 ⎦, …
0 0 a5
so
⎡ ⎤ ⎡ ⎤ ⎡ 2 ⎤ ⎡ 3 ⎤
1 0 0 a 1 0 a 2a 1 a 3a 2 3a
1 1
eJ x = ⎣0 1 0⎦ + ⎣0 a 1⎦ x + ⎣ 0 a 2 2a ⎦ x 2 + ⎣ 0 a 3 3a 2 ⎦ x 3
2! 3!
0 0 1 0 0 a 0 0 a2 0 0 a3
⎡ 4 ⎤ ⎡ ⎤
a 4a 3 6a 2 a 5 5a 4 10a 3
1 ⎣ 1
+ 0 a 4 4a 3 ⎦ x + ⎣ 0 a 5
4
5a 4 ⎦ x 5 + . . .
4! 5!
0 0 a4 0 0 a5
⎡ ⎤
m11 m12 m13
=⎣ 0 m22 m23 ⎦ ,
0 0 m33
and we see that
1 1 1 1
m11 = m22 = m33 = 1 + ax + (ax)2 + (ax)3 + (ax)4 + (ax)5 + . . .
2! 3! 4! 5!
= eax ,
and
1 2 3 1 1
m12 = m23 = x + ax 2 + a x + a3x 4 + a4x 5 + . . .
2! 3! 4!
= xeax
as we saw in the 2-by-2 case. Finally,
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
and similarly, for larger Jordan blocks. 2
Let us see how to apply this theorem in a couple of examples.
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

−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
and so
Y = eAx

Ax γ1
=e
γ2

γ1 e2x + (−2γ1 + γ2 )xe2x
= .
γ2 e2x + (−4γ1 + 2γ2 )xe2x
The initial value problem has solution
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
⎡ ⎤ ⎡ ⎤
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
⎡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
The initial value problem has solution
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.
Example 4.12. (Compare Example 2.4.) Consider the system

2 −17
Y = AY where A = .
1 4
We saw in Example 2.4 that A = P J P −1 with

−1 + 4i −1 − 4i 3 + 4i 0
P = and J = .
1 1 0 3 − 4i
Then
eAx = P eJ x P −1
−1
−1 + 4i −1 − 4i e(3+4i)x 0 −1 + 4i −1 − 4i
=
1 1 0 e(3−4i)x 1 1
(3+4i)x
−1 + 4i −1 − 4i e 0 1 1 + 4i
= (1/(8i))
1 1 0 e(3−4i)x −1 −1 + 4i

m11 m12
=
m21 m22
where
m11 = (1/(8i))((−1 + 4i)e(3+4i)x + (−1 − 4i)(−e(3−4i)x ))
= (1/(8i))((−1 + 4i)e3x (cos(4x) + i sin(4x)) − (−1 − 4i)e3x (cos(4x) − i sin(4x))
= (1/(8i))(ie3x (4 cos(4x) − sin(4x))(2)) = e3x (cos(4x) − (1/4) sin(4x)),
m12 = (1/(8i))((−1 + 4i)(1 + 4i)e(3+4i)x + (−1 − 4i)(−1 + 4i)e(3−4i)x )
= (1/(8i))(ie3x (−17 sin(4x))(2)) = e3x ((−17/4) sin(4x)),
m21 = (1/(8i))(e(3+4i)x − e(3−4i)x )
= (1/(8i))(ie3x (sin(4x))(2)) = e3x ((1/4) sin(4x)),
m22 = (1/(8i))((1 + 4i)e(3+4i)x + (−1 + 4i)e(3−4i)x )
= (1/(8i))((1 + 4i)e3x (cos(4x) + i sin(4x)) + (−1 + 4i)e3x (cos(4x) − i sin(4x))
= (1/(8i))(ie3x (4 cos(4x) + sin(4x))(2)) = e3x (cos(4x) + (1/4) sin(4x)) .
Thus,
−17
cos(4x) − 41 sin(4x) 4 sin(4x)
eAx = e3x
1
4 sin(4x) cos(4x) + 41 sin(4x)
and
γ1 e3x cos(4x) + ( −1
4 γ1 +
−17 3x
4 γ2 )e sin(4x)
Y =e Ax
= .
γ2 e3x cos(4x) + ( 41 γ1 + 41 γ2 )e3x sin(4x)
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 .
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).
Thus, consider Y = AY + G(x). Then, proceeding analogously as for an ordinary first-order

linear differential equation, we have
Y = AY + G(x)

Y − AY = G(x)
and, multiplying this equation by the integrating factor e−Ax , we obtain
e−Ax (Y − AY ) = e−Ax G(x)

(e−Ax Y ) = e−Ax 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.
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

In each exercise:
(a) Find eAx and the solution Y = eAx of Y = AY .
(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.
This theorem leads to the following definition.
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
is the coordinate vector of v in the basis B .
Remark 1.3. In particular, we may take V = Cn and consider the standard basis E = {e1 , . . . , en }
where
⎡ ⎤
0
⎢ .. ⎥
⎢.⎥
⎢ ⎥
⎢0⎥
ei = ⎢
⎢1⎥ ,
⎥
⎢ ⎥
⎢0⎥
⎣ ⎦
..
.
with 1 in the i th position, and 0 elsewhere.

70 APPENDIX A. BACKGROUND RESULTS
Then, if
⎡ ⎤
c1
⎢ ⎥ c2
⎢ ⎥
⎢ ⎥ ..
v=⎢ ⎥ , .
⎢ ⎥
⎣cn−1 ⎦
cn
we see that ⎡ ⎤ ⎡ ⎤
1 0
⎢0⎥ ⎢0⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
v = c1 ⎢ ... ⎥ + . . . + cn ⎢ ... ⎥ = c1 e1 + . . . + cn en
⎢ ⎥ ⎢ ⎥
⎣0⎦ ⎣0⎦
0 1
so we then see that ⎡ ⎤
c1
⎢ ⎥
c2
⎢ ⎥
⎢ ⎥..
[v]E = ⎢ ⎥ .
.
⎢ ⎥
⎣cn−1 ⎦
cn
(In other words, a vector in Cn “looks like” itself in the standard basis.)
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 .
Furthermore, the matrix [T ]B is given by

[T ]B = [v1 ]B [v2 ]B . . .
[vn ]B .

Similarly, this theorem leads to the following definition.
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 .
This matrix has the following properties:

(1) For any vector v in V ,
[v]C = PC ←B [v]B .
(2) This matrix is invertible and

(PC ←B ) −1
= PB←C = [w1 ]B [w2 ]B . . .
[wn ]B .

(3) For any linear transformation T : V −→ V ,
[T ]C = PC ←B [T ]B PB←C
= PC ←B [T ]B (PC ←B )−1
= (PB←C )−1 [T ]B PB←C .
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 .
Now we come to what is the crucial point for us.
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 .

Proof. By Theorem 1.7,

[TA ]E = PE ←B [TA ]B (PE ←B )−1 .
But by Remark 1.3,

PE ←B = [v1 ]E [v2 ]E . . .
[vn ]E = v1 v2 . . .

vn = P ,

and by Remark 1.6,

[TA ]E = A .
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

is a matrix whose columns are generalized eigenvectors forming a chain.
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 ,
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
Lemma 1.12. Let a be an eigenvalue of a matrix A. Then
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
A.2 PROPERTIES OF THE COMPLEX EXPONENTIAL

In this section of the Appendix, we prove properties of the complex exponential. For convenience,
we restate the basic definition and the properties we are trying to prove.
Definition 2.1. For a complex number z, the exponential ez is defined by

ez = 1 + z + z2 /2! + z3 /3! + . . . .
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.
Theorem 2.2. (1) (Euler) For any θ,

eiθ = cos(θ ) + i sin(θ ) .
(2) For any a,
d az
(e ) = aeaz .
dz
(3) For any z1 and z2 ,
ez1 +z2 = ez1 ez2 .
(4) If z = s + it, then
ez = es (cos(t) + i sin(t)) .
(5) For any z,
ez = ez .
Proof. (1) We begin with the following observation: i 0 = 1, i 1 = i, i 2 = −1, i 3 = i 2 i = −i,

i 4 = i 3 i = 1, i 5 = i 4 i = i, i 6 = i 4 i 2 = −1, i 7 = i 4 i 3 = −i, etc. In other words. the powers of
i, beginning with i 0 , successively cycle through 1, i, −1, and −i. With this in mind, we compute
directly from the definition:
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:
eiθ = (1 − θ 2 /2! + θ 4 /4! − θ 6 /6! + . . .) + i(θ − θ 3 /3! + θ 5 /5! − θ 7 /7! + . . .) .
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
ec+dz = 1 + (c + dz) + (c + dz)2 /2! + (c + dz)3 /3! + (c + dz)4 /4! + . . .

c+dz
(e ) = d + 2d(c + dz)/2! + 3d(c + dz)2 /3! + 4d(c + dz)3 /4! + . . .
= d + d(c + dz) + d(c + dz)2 /2! + d(c + dz)3 /3! + . . .
= d(1 + (c + dz) + (c + dz)2 /2! + (c + dz)3 /3! + . . .)
= dec+dz .
Now, setting c = 0 and d = a, we obtain (2).

A.2. PROPERTIES OF THE COMPLEX EXPONENTIAL 77
(3) Setting c = a and d = 1 in the above formula, we find that (ea+z ) = ea+z . In other words,
f1 (z) = e a+z
is a solution of the differential equation f (z) = f (z), and f1 (0) = ea+0 = ea . On the
other hand, setting f2 (z) = ea ez , we see from (2), and the fact that ea is a constant, that f2 (z) is also
a solution of the differential equation f (z) = f (z), and f2 (0) = ea e0 = ea . Thus, f1 (z) and f2 (z)
are solutions of the same first-order linear differential equation satisfying the same initial condition,
so by the fundamental existence and uniqueness theorem (also valid for complex functions), they
must be equal. Thus, ea+z = ea ez . Setting a = z1 and z = z2 , we obtain (3).
(4) This follows directly from (1) and (3):
ez = es+it = es eit = es (cos(t) + i sin(t)) .
(5) This follows directly from (1) and (4). Let z = s + it, so z = s − it. We compute:
ez = es−it = es+i(−t) = es (cos(−t) + i sin(−t)) = es (cos(t) + i(− sin(t))

= es (cos(t) − i sin(t)) = es (cos(t) + i sin(t)) = ez .
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

−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 = .

γ1 e3x + (−21γ1 + 9γ2 )xe3x
Y = .
γ2 e3x + (−49γ1 + 21γ2 )xe3x

e7x − 5xe7x xe7x
5a. eAx = .

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
⎡ ⎤
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 = .

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 = .
85
Index
antiderivative general solution of homogenous system

arbitrary, 48, 67 of, 25, 26, 29, 36, 43, 47, 54
general solution of inhomogenous system
chain of generalized eigenvectors, 10, 12, 73 of, 47
bottom of, 10 homogeneous system of, 25, 29, 36, 40, 54
top of, 10 inhomogeneous system of, 46, 67
characteristic polynomial, 1, 75 initial value problem for system of, 35, 36,
complex root, see eigenvalue, complex 52, 54
complex exponential, 41, 75 particular solution of inhomogeneous
diagonalizable, 5–7, 72 system of, 48, 67
uncoupled system of, 27, 31
eigenspace, 1 linear transformation, 70
generalized, 8, 9 matrix of, 70
eigenvalue, 1
complex, 40, 43, 65 matrix
eigenvector, 1 change-of-basis, 72
generalized, 8 matrix exponential, 38, 48, 53–55, 60, 61,
index of generalized, 8 66–68
Euler, 41, 75 multiplicity, 2
algebraic, 2, 3, 5, 9, 72, 74
Fundamental Theorem of Algebra, 3 geometric, 2, 3, 5, 17, 72, 74
integrating factor, 30, 31, 67
similar, 4, 5
JCF, see Jordan Canonical Form standard basis, 69–71
Jordan block, 7, 8, 12, 28, 47, 60, 61, 68, 73
vector
Jordan Canonical Form, 8, 14, 17, 21, 25, 31,
coordinate, 69
47, 66
imaginary part of, 43
linear differential equations real part of, 43
associated homogenous system of, 47 vector space
fundamental matrix of system of, 26 basis of, 69

Jordan Canonical Form - Application To Differential Equations

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Jordan Canonical Form - Application To Differential Equations

Загружено:

Авторское право:

Доступные форматы

Jordan Canonical Form:

Jordan Canonical Form: Application to Differential Equations

ISBN: 9781598298048 paperback

A Publication in the Morgan & Claypool Publishers series

SYNTHESIS LECTURES ON MATHEMATICS AND STATISTICS #2

1 Jordan Canonical Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Solving Systems of Linear Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

B Answers to Odd-Numbered Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

As is customary in textbooks, we provide the answers to the odd-numbered exercises here.

Deﬁnition 1.1. If v = 0 is a vector such that, for some λ,

then v is an eigenvector of A associated to the eigenvalue λ.

We note that the deﬁnition of an eigenvalue/eigenvector can be expressed in an alternate form.

For an eigenvalue λ of A, we let Eλ denote the eigenspace of λ,

(The kernel Ker(A − λI ) is also known as the nullspace NS(A − λI ).)

We now introduce two important quantities associated to an eigenvalue of a matrix A.

Lemma 1.7. Let a be an eigenvalue of a matrix A. Then

Proof. In this case, applying Lemma 1.7, we have

Lemma 1.10. Let A be an n-by-n matrix. The following are equivalent:

Proof. (1) Suppose we have a linear combination 0 = c1 v1 + c2 v2 + . . . + cm vm . We need to show

0 = (A − am I )(c1 v1 + c2 v2 + . . . + cm−2 vm−2 + cm−1 vm−1 + cm vm )

0 = (A − am−1 I )(c1 (a1 − am )v1 + c2 (a2 − am )v2 + . . .

0 = c1 (a1 − a2 ) · · · (a1 − am−1 )(a1 − am )v1 .

But v1 = 0, as by deﬁnition an eigenvector is nonzero. Also, the product (a1 − a2 ) · · · (a1 −

0 = c1,1 v1,1 + . . . + c1,i1 v1,i1 + c2,1 v2,1 + . . . + c2,i2 v2,i2

0 = (A − a2 I )0 = (A − a2 I )(v1 + v2 ) = (A − a2 I )v1 + (A − a2 I )v2

so 0 = v1 ; similarly, 0 = v2 . But 0 = v1 = c1,1 v1,1 + . . . + c1,i1 v1,i1 implies c1,1 = . . . c1,i1 = 0,

Here is the main result of this section.

First let us suppose that for each eigenvalue a of A, geom-mult(a) = alg-mult(a).

and we see that ji , the i th column of J , is the vector

with ai in the i th position, and 0 elsewhere.

so the i th column of AP is Avi . But

and the i th column of P J is P ji ,

Remembering what the vector ji is, and multiplying, we see that

completing the proof of this half of the Theorem.

1.2 THE GENERAL CASE

with each Ji a Jordan block.

Deﬁnition 2.4. If v = 0 is a vector such that, for some λ,

Let us note that if v is a generalized eigenvector of index 1, then

and so v is an (ordinary) eigenvector.

We let E˜λ denote the generalized eigenspace of λ,

E˜λ = {v | (A − λI )k (v) = 0 for some k} .

It is easy to check that E˜λ is a subspace.

Let us next consider a generalized eigenvector vk of index k associated to an eigenvalue λ, and

Deﬁnition 2.7. If {v1 , . . . , vk } is a set of generalized eigenvectors associated to the eigenvalue λ of

vk−1 =(A − λI )vk , vk−2 = (A − λI )vk−1 , vk−3 = (A − λI )vk−2 ,

Proof. Suppose we have a linear combination

We must show each ci = 0.

c1 (A − λI )k−1 vk + c2 (A − λI )k−2 vk + · · · + ck−1 (A − λI )vk + ck vk = 0 .

c1 (A − λI )2k−2 vk + c2 (A − λI )2k−3 vk + · · · + ck−1 (A − λI )k vk + ck (A − λI )k−1 vk = 0 .

Now (A − λI )k−1 vk = v1 = 0. However, (A − λI )k vk = 0, and then also (A − λI )k+1 vk =

Since v1 = 0, this shows ck = 0, so our linear combination is

so {v1 , . . . , vk } is linearly independent. 2

is a matrix whose columns are generalized eigenvectors forming a chain.

and we see that ji , the i th column of J , is the vector

with 1 in the (i − 1)st position, a in the i th position, and 0 elsewhere.

so the i th column of AP is Avi . But

On the other hand,   

and the i th column of P J is P ji ,

Remembering what the vector ji is, and multiplying, we see that

P ji = vi−1 + avi for i > 1, = avi for i = 1

(iii) A has a single eigenvalue a of algebraic multiplicity 3.

(a) A has three linearly independent eigenvectors, u, v, and w. Then A = P J P −1 with

On the other hand,