You are on page 1of 7

Examples of how to use the Cayley-Hamilton

theorem to solve simultaneous equations when A


cannot be diagonalized
Solve

x = x y 6z
y = x + 3z
z = y + z.

This is equivalent to solving the matrix differential equation

X = AX, (1)

x x 1 1 6
where X = y , X = y and A = 1 0 3.
z z 0 1 1
Given the heading of this page, you might guess that A cannot be diagonalized,
and youd be right. Im going to find the eigenvectors of A to confirm that A
cant be diagonalized (and also because in this and many other examples, they
turn out to be useful later).
We get the eigenvalues of A by factorizing

CA () = det(A I3 )
= 3 + 3 2
= (1 )2 (2 ).

The eigenvector corresponding to = 2 is found by solving



1 1 6 : 0
1 2 3 : 0 ,
0 1 3 : 0

which after a couple of row operations gives



1 0 3 : 0
0 0 0 : 0 ,
0 1 3 : 0

3
meaning that x = 3z and y = 3z so that the eigenvector is V1 = c1 3.
1
For = 1 its a similar procedure: we reduce

2 1 6 : 0
1 1 3 : 0 ,
0 1 0 : 0

1
which after a couple of row operations gives

1 0 3 : 0
0 1 0 : 0 ,
0 0 0 : 0

3
meaning that x = 3z and y = 0, so that V2 = c2 0.
1
We have no further eigenvalues, so there will not be a third independent eigen-
vector. Therefore we cannot construct the matrices T or T 1 , and so A cannot
be diagonalized.
Now for the part that you came here to see.
We note two things:

1. If D = t is the usual D-operator then from (1), we have

DX = AX,

and its easy to see that this means that

D n X = An X

for any positive integer n. (Exercise: convince yourself of this.)


2. The Cayley-Hamilton theorem guarantees that A satisfies its own charac-
teristic equation, which in this case means that we are guaranteed that

A3 + 3A 2I3 = 0,

or A3 3A + 2I3 = 0 if you prefer.

Putting these two observations together, we can say that

(D3 3D + 2)X = D3 X 3DX + 2X


= A3 X 3AX + 2X
= (A3 3A + 2I3 )X
= 0X
= 0,

i.e. we have (D3 3D + 2)X = 0.


This is an easy D-operator equation to solve: you just use the same techniques
that you learned in Calc Ch.1, but remember that X here is a vector, so the
arbitrary constants we find will also be vectors.
So we solve (D 1)2 (D + 2)X = 0 to get

X = (P + Qt)et + Re2t

(dont use the name A because thats the name of our matrix). Here P , Q and
R are 3-vectors (vectors with 3 components), so there are actually 9 constants

2
hidden inside the solution we just found. 9 seems like quite a lot for a 3rd-order
differential equation (or for a system of 3 first-order equations), but we can
eliminate some of these constants by expressing them in terms of the others.
Heres how.
Notice that since X = (P + Qt)et + Re2t = P (et ) + Q(tet ) + R(e2t ), we can
write

X = (P + Qt)et + Qet 2Re2t


= (P + Q)et + Qtet 2re2t .

We also know that X = AX, so we can write

(P + Q)et + Qtet 2Re2t = AP et + AQtet + ARe2t . (2)

The thing about this equation is that its true for all values of t - in other words
its an identity, and what it says is that two combinations of functions are equal
to each other. Here, the functions of t are et , tet and e2t , which are linearly
independent (check this by computing the Wronskian if you dont believe me).
The only way this is possible is if the coefficients of these functions are equal,
so in (2) we can equate the coefficients of et , tet and e2t separately as follows:

P + Q = AP (3)
Q = AQ (4)
2R = AR. (5)

This is a system of 3 matrix equations, and in this example we have a little bit
of luck. Equation (4) says that Q is an eigenvector of A with eigenvalue = 1.
Similarly, (5) says that R is an eigenvector of A with eigenvalue = 2. With
my incredible foresight, I calculated those eigenvectors early on in this example,
so we can write down the solutions to (4) and (5) right away:

3 3
Q = q 0 and R = r 3 ,
1 1

where q and r are arbitrary constants.


Now we just have to solve (3), which we do by rewriting as (A I3 )P = Q and
then constructing and reducing an augmented matrix:

2 1 6 : 3q
1 1 3 : 0 ,
0 1 0 : q

which (eventually) gives


1 0 3 : q
0 1 0 : q ,
0 0 0 : 0

3

p1
i.e. if P = p2 then
p

p1 = q 3p
p2 = q
p = p,

giving
3 1
P = p 0 + q 1 .
1 0

So now we can finally write down the solution to our original system of equations:

X = (P + Qt)et + Re2t

3 1 3 3
= p 0 + q 1 + q 0 t et + r 3 e2t ,
1 0 1 1

which means that

x = (3p + q 3qt)et + 3re2t


y = qet 3re2t
z = (p + qt)et + re2t .

We managed to eliminate all but 3 arbitrary constants, which is the right number
to have in the solution of a system of 3 first-order equations, and were done.

4
Another example
Solve
x = 3x + 3z
y = 2x y 3z
z = x + y + 3z
which amounts to solving

3 0 3
X = 2 1 3 X.
1 1 2

First thing, weregoing to need


to find the characteristic polynomial CB () of
3 0 3
the matrix B = 2 1 3. Instead of calculating it directly, this time
1 1 2
Ill show you how using row operations can give you the polynomial already
factorized.

3 0 3
CB () = det 2 1 3
1 1 2

3 0 3 R1
= det 1 1 0 R1 + R2
1 1 2 R3

3 0 3
= (1 + ) det 1 1 0
1 1 2

3 0 3 R1
= (1 + ) det 1 1 0 R2
2 0 2 R3 R2 because 0s make life easier
= (1 + )[(3 )(2 ) + 6]
= (1 + )(2 + 6 + 6)
= (1 + )2 ,
so the eigenvalues are -1 and 0.
For = 1, we solve
2 0 3 : 0
2 0 3 : 0 ,
1 1 3 : 0
which after a couple of row operations gives
1 0 23 : 0

0 0 0 : 0 ,
3
0 1 2 : 0

3
giving eigenvector c1 3.
2

5
For = 0, solve
3 0 3 : 0
2 1 3 : 0 ,
1 1 2 : 0
which after a couple of row operations gives

0 0 0 : 0
0 1 1 : 0 ,
1 0 1 : 0

1
giving eigenvector c2 1.
1
To diagonalize B, wed need a third independent eigenvector, so B is not diag-
onalizable.
Instead we use the Cayley-Hamilton theorem, and in a way similar to the pre-
vious example, we say
(1 + D)2 DX = 0,
which has solution

X = (P + Qt)et + R
= P et + Qtet + R,

where, as before, P , Q and R are vectors with 3 components each. Now we


compute X:

X = P et + Qet Qtet
= (P + Q)et Qtet , (6)

which we know must be equal to

BX = BP et + BQtet + BR. (7)

We equate the coefficients of 1, et and tet in (6) and (7) to get the system of
equations

BP = P + Q (8)
BQ = Q (9)
BR = 0. (10)

Equations (9) and (10) say that Q and R are eigenvectors of B; Q is an eigen-
vector with eigenvalue = 1 and R is an eigenvector with eigenvalue = 0.
We worked out those eigenvectors already, so we can write

3 1
Q = q 3 and R = r 1 .
2 1

6
We just need to find P , which we do by rewriting (8) as (B + I3 )P = Q and
solving.
2 0 3 : 3q
2 0 3 : 3q ,
1 1 3 : 2q
which (eventually) gives

32 : 23 q

1 0
0 0 0 : 0 .
3 1
0 1 2 : 2q

(Notice that the matrix is exactly the same as the one we had to deal with when
finding the first eigenvector, its only the RHS that is different. So we do the
exact same row operations, but we only have to work out what happens on the
RHS because the LHS will be the same as before.)

p1
If P = p2 then
p

p1 = 32 q + 32 p
p2 = 12 q 32 p
p = p,

giving
3 3
P = 21 p 3 + 12 q 1 ,
2 0
and so our final solution is

x 3 3 3 1
y = 1 p 3 + 1 q 1 + qt 3 et + r 1 ,
2 2
z 2 0 2 1

where p, q and r are arbitrary (scalar) constants.

You might also like