Avoiding the Jordan Canonical Form in the Discussion of Linear Systems with Constant

Coefficients
Author(s): E. J. Putzer
Source: The American Mathematical Monthly, Vol. 73, No. 1 (Jan., 1966), pp. 2-7
Published by: Mathematical Association of America
Stable URL: http://www.jstor.org/stable/2313914
Accessed: 18-10-2017 10:13 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

Mathematical Association of America is collaborating with JSTOR to digitize, preserve and

extend access to The American Mathematical Monthly

This content downloaded from 143.107.45.11 on Wed, 18 Oct 2017 10:13:08 UTC
All use subject to http://about.jstor.org/terms
2 LINEAR SYSTEMS WITH CONSTANT COEFFICIENTS [January

In addition to serving ex officio on the Board of Governors and various

Association committees (the Executive and Finance Committees, the Commit-
tees on Publications and on Membership, and the Joint Committee on Places
of Meetings), he has served by appointment on nominating committees, on the
planning committee for the fiftieth anniversary celebration, and as chairman of
the Committee on the Structure of the Government of the Association. He has
represented the Association in the Division of Mathematics of the National
Research Council, on the Policy Committee for Mathematics and then the Con-
ference Board of the Mathematical Sciences, where he was elected to an addi-
tional term as member-at-large. He was Chairman in 1951-52 of the rnathe-
matics section of the American Society for Engineering Education and partici-
pated three summers (1959-61) in writing projects of the School Mathematics
Study Group.
Over and beyond this handsome record of accomplishment shine the amiable
warmth and sterling spirit of the man himself-and his wife, Marian, who has
shared in his mission to mathematics. It has been service with the extra zeal and
zest of a man who enjoys working with other people and accepts their foibles
with a chuckle. For all this we proudly and gratefully hail Harry Merrill
Gehman.
A. W. TUCKER

AVOIDING THE JORDAN CANONICAL FORM IN THE

DISCUSSION OF LINEAR SYSTEMS WITH
CONSTANT COEFFICIENTS

E. J. PUTZER, North American Aviation Science Center, Thousand Oaks, California

(Now at 5143 Topanga Canyon, Woodland Hills, California)

Consider the differential equation

(1) x=Ax; x(O)=xo; 0 t <oo,

where x and x0 are n-vectors and A is an n X n matrix of constants. In this paper

we present two methods, believed to be new, for explicitly writing down the
solution of (1) without making any preliminary transformations. ThWs is par-
ticularly useful, both for teaching and applied work, when the matrix A cannot
be diagonalized, since the necessity of discussing or finding the Jordan Canonical
Form (J.C.F.) of A is completely by-passed.
If eAt is defined as usual by a power series it is well known (see [1]) that the
solution of (1) is

x = eAtxo,

This content downloaded from 143.107.45.11 on Wed, 18 Oct 2017 10:13:08 UTC
All use subject to http://about.jstor.org/terms
1966] LINEAR SYSTEMS WITH CONSTANT COEFFICIENTS 3

so the problem is to calculate the function eAt. In [2], this is done via the J.C.F.
of A. In [1], it is shown how the J.C.F. can be by-passed by a transformation
which reduces A to a triangular form in which the off diagonal elements are
arbitrarily small. While this approach permits a theoretical discussion of the
form of exp {At} and its behavior as t becomes infinite (Theorem 7 of [1]), it
is not intended as a practical method for calculating the function. The following
two theorems suggest an alternate approach which can be used both for cal-
culation and for expository discussion. It miiay be noted that the formula of
Theorem 2 is simpler than that of Theorem 1 since the ri are easier to calculate
than the qj.
In order to state Theorem 1 simply, it will be convenient to introduce some
notation.
Let A be an n Xn matrix of constants, and let

f(X)| XI- A |XEn + C_Xn-' + + C1X + Co

be the characteristic polynomial of A. Construct the scalar function z(t) which

is the solution of the differential equation

(2) z + c,_1Z`(' ) + + cli + coz 0

with initial conditions

(3) z(0) = *(0) Z (-2) (0) = 0; z(7 ')(0) = 1.

We observe at this point that regardless of the multiplicities of the roots of

f(X) =0, once these are obtained it is trivial to write down the general solution
of (2). Then oine solves a single set of linear algebraic equations to satisfy the
initial conditions (3). Since the right member of each of these equations is zero
except for the last, solving them entails only finding the cofactors of the ele-
ments of the last row of the associated matrix. It is not necessary to invert the
matrix itself. For teaching purposes, the point is that the form of the general
solution of (2) can be obtained quickly and easily by elementary methods.
Now define

[ Z() 1 Ci C2 * * n- 1

=
Z(z'=
Lz(t'(tl'and
0 (t) ], C3 . .1
C=
L (n-l) (t,) Cn-1 1

We then have the following

THEOREM 1.

n-I

(4) eAt = qE q,(I)A ,

j=O

This content downloaded from 143.107.45.11 on Wed, 18 Oct 2017 10:13:08 UTC
All use subject to http://about.jstor.org/terms
4 LINEAR SYSTEMS WITH CONSTANT COEFFICIENTS [January

where qo(t), * q".-(t) are the elements of the column vector

(5) q(t) = CZ(t).

Before we prove this, a remark is in order as to what happens when f(X) has
multiple roots but the minimal polynomial of A has distinct factors so that A
can in fact be diagonalized. It appears at first glance that our formula will con-
tain powers of t, yet we know this cannot be the case. What occurs, of course, is
that the powers of t in (4) just cancel each other out. This is the nice feature of
formula (4); it is true for all matrices A, and we never have to concern ourselves
about the nature of the minimal polynomial of A, or its J.C.F., and no pre-
liminary transformations of any kind need be made.
Proof. We will show that if
n-I

(6) E(t)-E qj(t) A

j=O

then dcI/dt = A4? and 4(O) = I, so 1(t) = eAt. Since only qo(t) involves z(n1)(t),
qj(O) = 0 for j > 1. Clearly, qo(O) = 1. Thus, b(0) = I.
Now consider (d4/dt) -A4. Differentiating (6), and applying the Hamilton-
Cayley Theorem
n-1

An + E cjAi = 0,
j=O

we obtain

d'I' n-1
d---A ?= (4o +

It will suffice, therefore, to show that

qO0(t) --coqn-1 (t),

qj (t)- i_l (t)- cjqn-l (t) j ,(n-1)

From the definition (5),

n-j-1

(7) qj(t) = E Ck+jZ(k') + z(n-l)

k-1

Therefore 4j(t) = j;-`-1 Ck+jZ(k) +Z(n-i).

But qn-l=z, so we have

n-Ci l
(8) + - + cqn+ z(n) for = 0, 1, * , n - 1.
k=O

If j=0, this yields

n-1

4o + coqn-I = ckz(k) + Z(n)

k=O

This content downloaded from 143.107.45.11 on Wed, 18 Oct 2017 10:13:08 UTC
All use subject to http://about.jstor.org/terms
1966] LINEAR SYSTEMS WITH CONSTANT COEFFICIENTS 5

which is zero because of (2).

If j > 1, replace j by j-1 in (7) and change the summation index from k to
k+1 to get
n-f-i

(9) qj-i(t) - Ck+jZ(l) + Z(n-j).

kaO

Comparing (9) and (8) we have 4j +cjqn_(t)- qj_.(t) forj= 1, 2, n-i.

Students will want to see the formula (4) derived. One merely presents the
proof backward, beginning with the observation that because of the Hamilton-
Cayley theorem, exp {At} should be expressible in the form (6). Regarding the
qj as unknowns, and applying the differential equation which exp {At} satis-
fies, leads directly to (4).
A second explicit formula for exp { At }, which also holds for all matrices A,
is the following: Let N1, N2, .. *Xn be the eigenvalues of A in some arbitrary but
specified order. These are not necessarily distinct. Then

THEOREM 2. eAt= - 0 rjl+(t)Pj, where

PO = I; P} I (A - kI)j ( j=1,**,n)
k=1

and r1(t), * * , r.(t) is the solution of the triangular system

f 1= X1r1

* = rj-1 + Xjr, (j = 2, ... n)

lri(0) = 1; rj(0) = 0 (j= 2,* . .n).
Proof. Let

n-1

(10) @(t)--E rj(t)P

j=0

and define ro(t) 0. Then from (10) and the equations satisfied by the ri(t) we
have, after collecting terms in rj,
n-2

n- A = E2 [Pj+l + (X,+1 - Xn)P3Irj?il

0

Using Pj,i (A - Xj+I)Pj in this gives

-Xnb- = (A - XnI)(I - rn(t)Pn-l)

= (A - XnI)4? - rn(t)Pn.

But Pn -0 from the Hamilton-Cayley Theorem, so = A. Then since b(0) I

it follows that 4(t) = eAt.
Example. If one desires a numerical example for class presentation an ap-

This content downloaded from 143.107.45.11 on Wed, 18 Oct 2017 10:13:08 UTC
All use subject to http://about.jstor.org/terms
6 LINEAR SYSTEMS WITH CONSTANT COEFFICIENTS [January

propriate matrix A can be prepared in advance by arbitrarily choosing a set of

eigenvalues, a Jordan Canonical Form J and a nonsingular matrix S and
calculating

A = SJS-1.

Then beginning with A, one simply calculates the set { qi(t) } and/or { ri(t) }.
sider the case of a 3 X 3 matrix having eigenvalues (X, X, ,). There are two sub
cases; the one in which the normal form of A is diagonal, and the one in which
it is not. These two subcases are taken care of automatically by the given formula
for exp {At}, and do not enter at all into the calculation of the {qi} or {ri}.
As an example we will explicitly find the sets { qi } and { ri } for the case of a
3 X 3 matrix with eigenvalues (X, X, X). We note that aside from the trivial case
in which the normal form (and hence A itself) is diagonal, there are two distinct
nondiagonal normal forms that A may have. As above, these do not have to be
treated separately.
From Theorem 1,

f(x) (x - X)3 = X- 3Xx2 + 3X2xI-3

so Cl = 3X2, c2 =- 3X. Obviously z(t) = (a,+a2t+a3t2)elt. Applying the initia

ditions to find the ai yields z = -t2ellt, so
[ t21
Z(t) = XeXt Xt2 + 2t
X2t2 + 4Xt + 2_

Then since

[ 3X2 -3x 1]

c= -3x 1 0
1 0 0

-X2t2 - 2Xt + 2-

q = CZ(t) =let -2Xt2+ 2t K

Thus

(11) eAt = 2ext{ (X2t2- 2Xt + 2)1 + (-2Xt2 + 2t)A + t2A2}

f or every 3 X 3 matrix A having all three eigenvalues equal to X. The correspond-

ing formula from Theorem 2 is obtained by solving the system

kI = 1rl

r2= r1 + Xr2

kt3 = r2+ Xr3

This content downloaded from 143.107.45.11 on Wed, 18 Oct 2017 10:13:08 UTC
All use subject to http://about.jstor.org/terms
 19661 ASYMPTOTIC ANALYSIS IN APPLIED MATHEMATICS 7

with the specified initial values. This immediately gives

ri = e7 ; r2 = le,Xf ; r3 = et2 CM

so

(12) eAt = 1ext{2I + 2t(A- XI) + t2(A- XI)2}.

Of course, if we collect powers of A in (12) we will obtain (11).

References

1. R. E. Bellman, Stability Theory of Differential Equations, McGraw-Hill, New York, 1953

2. E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equations, McGraw-
Hill, New York, 1955.

THE IMPORTANCE OF ASYMPTOTIC ANALYSIS

IN APPLIED MATHEMATICS

L. A. SEGEL, Rensselaer Polytechnic Institute

Friedrichs began his 1954 Gibbs lecture on asymptotic phenomena in mathe-

matical physics [6] by saying "the problems I intend to speak about belong to
the somewhat undefined and disputed region at the border between mathematics
and physics." In the ten years since his lecture there has been a growing ten-
dency to define this region and some others as the domain of applied mathematics.
A number of explicit university applied mathematics programs have been
started, the Society for Industrial and Applied Mathematics has been founded
in the United States and the Institute of Mathematics and its Applications in
England.
It ought to be helpful to have on record various opinions on what this new,
or renewed, activity is all about. The general nature of applied mathematics has
been ably discussed [7], [8] and the time now appears ripe for more detailed
remarks on the elements of an applied mathematical outlook. The purpose of
the following supplement to Friedrichs' remarks and references is to stress the
importance of asymptotic analysis in applied mathematics. Although they will
not be mentioned here, there are of course several other important aspects of
the subject. One of these is the use of computing machines. Murray [13] con-
centrates so exclusively on this one aspect that he gives the impression that
applied mathematics is a problem solving service rather than the independent
science envisioned here and elsewhere. An indication of the central position of
asymptotics is that in concentrating on it we do not present an unbalanced
picture of applied mathematics.

This content downloaded from 143.107.45.11 on Wed, 18 Oct 2017 10:13:08 UTC
All use subject to http://about.jstor.org/terms