You are on page 1of 7

Linear Algebra Notes

Chapter 9

MULTIPLE EIGENVALUES AND NILPOTENT MATRICES

Before taking up multiple eigenvalues, we begin with a


Prelude: Take a matrix A, and an invertible matrix B. We get a new matrix
B 1 AB, which is called conjugate to A. We have seen that determinant and
trace have the conjugation property:

det(B 1 AB) = det(A), tr(B 1 AB) = tr(A).

Since the characteristic polynomial is made out of the trace and determinant, it
follows that A and B 1 AB have the same characteristic polynomials:

PA (x) = PB 1 AB (x).

This means A and B 1 AB have the same eigenvalues. They usually do not have
the same eigenvectors, but their eigenvectors are related: If Au = u, then

B 1 AB(B 1 u) = B 1 Au = B 1 u,

so B 1 u is an eigenvector for B 1 AB with eigenvalue . In summary,


Proposition 1. Conjugate matrices A and B 1 AB have the same trace and de-
terminant, the same characteristic polynomial, and the same eigenvalues. If u is a
-eigenvector for A, then B 1 u is a -eigenvector for B 1 AB.
end of Prelude.

Now to business. We have seen that if A has distinct real eigenvalues , , then
we can find B so that  
1 0
B AB = . (9a)
0
By Proposition 1, the diagonal matrix in (9a) shares many properties with A. How-
ever, being diagonal, it is for many purposes simpler than A. We seek an equation
like (9a) in the remaining cases: Multiple eigenvalues, and Complex eigenvalues.
In this chapter we consider multiple eigenvalues. We will always assume that the
entries of A are real.
We get multiple eigenvalues when

PA (x) = (x )2 = x2 2x + 2 .

This means
2 = tr(A) and 2 = det(A),
1
2

so
4 det(A) = [tr(A)]2 (9b)
The matrices with only one eigenvalue are those whose trace and determinant satisfy
(9b). Note that = 21 tr(A) will be real, since the entries of A are real.
There are then two possibilities for A. Either A = I, or A 6= I. Assume
A 6= I. Then, in contrast to previous situations, we cannot conjugate A to a
diagonal matrix. But we can do almost as well:
Proposition 2. If A has only one eigenvalue , and A 6= I, then there is an
invertible matrix B such that
 
1 1
B AB = .
0

The difference from the distinct-eigenvalue case is the 1 in the upper right entry,
which cannot be made zero, no matter how we choose B.
Proof. We give a recipe for finding B. As before, first find a -eigenvector u, for
A. Now take any vector v which is not proportional to u. Let B1 be the matrix
defined by
B1 e1 = u, B1 e2 = v.
We do a familiar computation

B11 AB1 e1 = B11 Au = B11 u = e1 .

So  
g
B11 AB1 = ,
0 h
for some numbers g, h that we dont yet know. But in fact, h = . To see this,
note that  
1 g
2 = tr(A) = tr(B1 AB1 ) = tr = + h,
0 h
(we used Proposition 1 at the second equality) so indeed, = h. So we in fact have
 
1 g
B1 AB1 = .
0

Now g 6= 0, since if g = 0 we would have A = I, but we are assuming A 6= I.


g 0
Now conjugate further using , and get
0 1
 1     
g 0 g g 0 1
= .
0 1 0 0 1 0

This shows that if we take  


g 0
B = B1 ,
0 1
then  
1 1
B AB = ,
0
3

as claimed in Proposition 2. 

Now if you want to compute An , you could first note that


 n  
1 n nn1
= (9b)
0 0 n

so    n 
1 1 n nn1
n
A = (B B ) =B B 1 . (9c)
0 0 n
Actually, there is a much better formula for An , but first lets have an example to
illustrate what weve seen so far.
Example:  
0 4
A= .
1 4

We find PA (x) = (x 2)2 so = 2 is a multiple eigenvalue of A. Our formula for


eigenvectors gives (b, a) = (4, 2). Scaling, we can take the eigenvector to be
u = (2, 1). Now we choose any another vector v which is not proportional to u.
Let us take v = e1 , so we have
 
2 1
B1 = .
1 0

We then compute  
0 1
B11 = ,
1 2
and      
0 1 0 4 2 1 2 1
B11 AB1 = = .
1 2 1 4 1 0 0 2
So g = 1 and we take
   
1 0 2 1
B = B1 = .
0 1 1 0

To check our calculations we compute


 
1 2 1
B AB = .
0 2
 
2 1
So B = does the job, as predicted by Proposition 2.
1 0
 n  n 
2 1 2 n2n1
Continuing further using equation (9c), we have = , so
0 2 0 2n
   n 
n 2n n2n1 1 2 n2n n2n+1
A =B n B = .
0 2 n2n1 2n + n2n
4

We could write this as a sum


 n   
n 2 0 n1 2 4
A = + n2 .
0 2n 1 2

The first matrix in this sum is the diagonal matrix whose diagonal entries are the
powers of the eigenvalue of A. The second matrix (considered without the factor
n2n1 ), is almost A, but with the eigenvalue 2 subtracted from the diagonal entries
of A. Let us call this matrix A0 . So
   
2 4 02 4
A0 = = = A 2I.
1 2 1 4 2

Thus, we could write our formula for An as

An = 2n I + n2n1 A0 .

This is a formula is much simpler than (9c), and it holds in general.


 
a b
Proposition 3. Let A = be a matrix with PA = (x )2 . Define the new
c d
matrix  
a b
A0 = .
c d
 
0 0
Then A20 = , and
0 0
An = n I + nn1 A0 .

The point of Proposition 3 is that it allows us to compute An in a much easier


way than equation (9c). We dont have to find B to use Proposition 3. However,
0 0
we will use B to prove Proposition 3. The key point is that A20 = .
0 0
Proof. We use our earlier formula
 
1
A=B B 1 ,
0

from Proposition 2. Note that


   
1 0 1
= I + ,
0 0 0
so
     
1 1 0 1 1 1 0 1
A=B B = B(I + )B = B(I)B + B B 1 .
0 0 0 0 0

Since I commutes with every matrix, we have B(I)B 1 = I, so


 
0 1
A = I + B B 1 .
0 0
5

But also A = I + A0 , so
 
0 1
I + A0 = A = I + B B 1 .
0 0

Subtracting I from both sides, we get


 
0 1
A0 = B B 1 .
0 0

Hence
   
0 1 1 0 1
A20 =B B B B 1
0 0 0 0
 2
0 1
=B B 1
0 0
 
0 0
=B B 1
0 0
 
0 0
= ,
0 0
 2  
0 1 0 0
because = . That was the key point. Now we have
0 0 0 0

An = (I + A0 )n .

Inside the parentheses, we have two matrices. One of them, I, commutes with all
matrices, and the other one, A0 , squares to zero.

Now let us pause, and temporarily empty our minds of all thoughts of A and A0 .
Clear the table. On this empty table, we put two matrices
 X,Y with the following
0 0
two properties. First, XY = Y X, and second, Y 2 = . Now we add them
0 0
and raise to powers. The second power is

(X + Y )2 = (X + Y )(X + Y ) = XX + XY + Y X + Y Y = X 2 + 2XY,

because XY = Y X and Y 2 = 0. In the general power

(X + Y )n = (X + Y )(X + Y ) (X + Y )

we get each term by choosing either X or Y in each factor. Since they commute,
we can put the Y s together, which will give zero unless we have at most one power
of Y . So we either choose X in each factor (one such term), or we choose Y in one
factor and X in all the others (n such terms). Thus

(X + Y )n = X n + nX n1 Y.

Now we resurrect A and A0 , and finish our calculation: Since (I)n = n I, we


get
An = (I + A0 )n = n I + nn1 A0 ,
6

as claimed. 
A matrix is called nilpotent if some power of it is zero. It turns out, for 2 2
matrices Y as we are considering here, that if Y n = 0, then already Y 2 = 0 (see
exercise 9.4). So a 22 matrix is nilpotent if either it is already zero, or if it squares
to zero.

We may summarize the main points in this chapter as follows. Any matrix A with
a multiple eigenvalue is a scalar matrix I plus a nilpotent matrix A0 = A I.
 A is not itself a scalar matrix, then A0 6= 0. In that case A is conjugate to
If
1 0 1
, and A0 is conjugate to . This last matrix does not involve . In
0 0 0
particular, all matrices A with multiple
 eigenvalues
 have their nilpotent parts A0
0 1
being conjugate to the same matrix .
0 0
Exercise 9.1. For each of thefollowing matrices A, find the eigenvalue , and a
1
matrix B such that B 1 AB = .
0
 
6 1
(a) A =
 1 8 
1 1
(b) A =
 1 1 
1 0
(c) A =
1 1 
0 1
(d) A =
1 2
(To make life easier for the grader, please choose v = e1 in each problem. Your
Bs will all be the same, having 1 in the upper right corner.)
 
1
Exercise 9.2. Show that every eigenvector of is proportional to e1 . (Hint:
0
Just calculate.)
Exercise 9.3. Let A be a matrix with multiple eigenvalue , and let u be an eigen-
vector of A. Assume that A is not a scalar matrix. Prove the following statements.
(a) Every eigenvector of A is proportional to u. (Hint: Proposition 1.)
(b) A0 u = (0, 0)
(c) If v is a vector for which A0 v = (0, 0), then v is an eigenvector of A.
(Taken together, these facts mean that A has exactly one eigenline, which consists
precisely of the vectors sent to (0, 0) by A0 . )
 
n 0 0
Exercise 9.4. Suppose A = for some n 0.
0 0
(a) What are the eigenvalues of A? (Hint: Apply A repeatedly to both sides of
the equation Au = u. )
(b) Compute A2 . (Hint: Use Proposition 3, and the value of from part (a). )

Exercise 9.5. Suppose = 0 is the only eigenvalue of A. Does this imply that A
is nilpotent?
7

Exercise 9.6. Show that the nilpotent matrices are precisely those of the form
 
a b
A= with a2 + bc = 0.
c a

(You have to show two things: First, that every nilpotent matrix has this form.
Second, that every matrix of this form is nilpotent.)