You are on page 1of 7

Diagonalization

1. Recap. If we have a discrete dynamical system ~x(t + 1) = A~x(t), where A is an n × n matrix, and
we’re given an initial condition ~x(0), what’s our strategy for finding ~x(t) in closed form?
Solution. ~x(t) = At ~x(0), but this is only easy to compute if there is an eigenbasis for A (i.e.,
a basis of Rn consisting of eigenvectors of A). If (~v1 , . . . , ~vn ) is an eigenbasis for A and λ1 , . . . , λn are
the corresponding eigenvalues, then our strategy was to express ~x(0) as a linear combination of the
eigenvectors, ~x(0) = c1~v1 + · · · + cn~vn . Then,

~x(t) = At ~x(0)
= c1 At~v1 + · · · + cn At~vn
= c1 λt1~v1 + · · · + cn λtn~vn

As we saw in class today, an equivalent way of thinking about this is that we are diagonalizing
 A;
λ1
that is, we are writing A = SDS −1 where S = ~v1 · · · ~vn and D = 
   .. , and then

.
λn
~x(t) = At ~x(0) = SDt S −1 ~x(0).

Note: When A does not have an eigenbasis (i.e., when A is not diagonalizable), we don’t have a simple
way of finding ~x(t).
 
−3
2. You are given the following information about a 3 × 3 matrix A: ~v1 =  1 is an eigenvector of A
  0  
1 0
with eigenvalue 3, ~v2 = 0 is an eigenvector of A with eigenvalue 2, and ~v3 = 2 is an eigenvector
5 7
of A with eigenvalue −2.

(a) Explain why B = (~v1 , ~v2 , ~v3 ) is a basis of R3 .

Solution. ~v1 , ~v2 , ~v3 must be linearly independent because they are eigenvectors of A with different
eigenvalues. Three linearly independent vectors in R3 must form a basis of R3 .

(b) If T (~x) = A~x, find the B-matrix of T .

Solution. We are told that T (~v1 ) = 3~v1 , T (~v2 ) = 2~v2 , and T (~v3 ) = −2~v3 . So,
     
3 0 0
[T (~v1 )]B = 0 , [T (~v2 )]B = 2 , [T (~v3 )]B =  0
0 0 −2
 
3
(Remember that saying [T (~v1 )]B = 0 means exactly that T (~v1 ) = 3~v1 + 0~v2 + 0~v3 .) Therefore,
0
 
3 0 0
the B-matrix of T is the diagonal matrix D = 0 2 0 .
0 0 −2

(c) How could you find a closed formula for A100 ?

1
Solution. If S = ~v1 ~v2 ~v3 , then we know A = SDS −1 . (This is just our usual formula for
 

converting between the standard matrix A of T and the B-matrix D of T .) Then,

A100 = (SDS −1 )100


= (SDS −1 )(SDS −1 )(SDS −1 ) · · · (SDS −1 )(SDS −1 )
| {z }
100 copies

We know that S −1 S = In , so we can rewrite this as:

S −1
= (SD )(S D−1
 S )( S −1
 S D ) · · · (S D−1
 S DS −1 )
 S )(
| {z }
100 copies

= SD100 S −1

What makes this worthwhile is that it’s actually very easy to compute D100 . Let’s start by
computing just a few powers of D:
    2 
3 0 0 3 0 0 3 0 0
D2 = 0 2 0  0 2 0  =  0 22 0 
0 0 −2 0 0 −2 0 0 (−2)2
 2    3 
3 0 0 3 0 0 3 0 0
D 3 = D 2 D =  0 22 0  0 2 0  =  0 23 0 
0 0 (−2)2 0 0 −2 0 0 (−2)3
 k 
3 0 0
The pattern should be pretty clear: Dk =  0 2k 0 . Therefore,
0 0 (−2)k

A100 = SD100 S −1
   100  −1
−3 1 0 3 −3 1 0
=  1 0 2  2100  1 0 2 ,
0 5 7 (−2)100 0 5 7

and this is something that we can multiply out directly. The point here is that it is much easier
to evaluate D100 than it is to evaluate A100 directly, because D is diagonal.
 
−8 10
3. Let A = .
−5 7

(a) A is diagonalizable; diagonalize it. That is, express A in the form SDS −1 where S is invertible
and D is diagonal.
Solution. The characteristic polynomial of A is λ2 + λ − 6 = (λ + 3)(λ
 − 2), so the eigenvalues
 
−5 10 2
of A are −3 and 2. The −3-eigenspace of A is ker(A + 3I) = ker = span . The
    −5 10 1 
−10 10 1 2 1
1-eigenspace of A is ker(A − 2I) = ker = span . So, A = SDS −1 with S =
−5 5 1 1 1
 
−3 0
and D = .
0 2

(b) Find At , where t is a positive integer; simplify the entries of the matrix At completely.

2
Solution. Since A = SDS −1 ,

At = SDt S −1
  t  −1
2 1 −3 0 2 1
=
1 1 0 2 1 1
t
   
2 1 (−3) 0 1 −1
=
1 1 0 2t −1 2
| {z }
S −1
2 1 (−3)t −(−3)t
  
=
1 1 −2t 2t+1
2(−3)t − 2t −2(−3)t + 2t+1
 
=
(−3)t − 2t −(−3)t + 2t+1

 
2
(c) Consider the discrete dynamical system ~x(t + 1) = A~x(t) with initial condition ~x(0) = .
1

Write a closed formula for ~x(t).


Solution. Since ~x(0) is an eigenvector of A with eigenvalue −3, ~x(t) = (−3)t ~x(0).

A2 A3 An
(d) Given that eA = I + A + + + ... + + .... (same as the power series of ex ; ex =
2! 3! n!
x2 x3 xn
1+x+ + + ... + + ....)
2! 3! n!
 λ 
−1 A D −1 D e 1 0
Let A = SDS . Prove that e = Se S . (Remark: if A is 2 by 2, then e = λ2 )
0 e

Solution. In class
Note: If A is diagonalizable with diagonal matrix B = S −1 AS one can define f (A) for any function f
by f (A) = Sf (B)S −1 where f is applied to each diagonal entry of B.

3
4. Decide whether each of the following matrices A is diagonalizable. If so, find a diagonal matrix D that
A is similar to; in addition, if A has 1 or 2 distinct eigenvalues, find an invertible matrix S such that
A = SDS −1 .
 
7 0 0
(a) A = 0 8 0
5 0 7
 
7−λ 0 0
Solution. The characteristic polynomial of A is det(A−λI3 ) = det  0 8−λ 0  = (7−
5 0 7−λ
λ)2 (8 − λ), so the eigenvalues are 7 with algebraic multiplicity 2 and 8 with algebraic multiplicity
1.
To decide whether A is diagonalizable, we really need to see whether the geometric multiplic-
ities of the eigenvalues add up to 3. Since 8 has algebraic multiplicity 1, it also has geo-
metric multiplicity1. To find the geometric multiplicity of 7, let’s look at the 7-eigenspace
0 0 0
ker(A − 7I3 ) = ker 0 1 0. We can see by inspection that A − 7I3 has rank 2, so its kernel is
5 0 0
1-dimensional. Therefore, the geometric multiplicity of 7 is 1, so A is not diagonalizable . (The
geometric multiplicities of the eigenvalues only add up to 2.)
 
1 2 −2
(b) A = 0 3 −2
0 0 1
 
1−λ 2 −2
Solution. The characteristic polynomial of A is det(A − λI3 ) = det  0 3−λ −2  =
0 0 1−λ
(1 − λ)2 (3 − λ), so the eigenvalues of A are 1 with algebraic multiplicity 2 and 3 with algebraic
multiplicity 1.
To decide whether A is diagonalizable, we really need to see whether the geometric multiplicities of
the eigenvalues add up to 3. Since 3 has algebraic multiplicity 1, it also has geometric multiplicity
1. To find the geometric multiplicity of 1, let’s look at the 1-eigenspace, or ker(A − I3 ) =
0 2 −2
ker 0 2 −2. We can see by inspection that A − I3 has rank 1, so its kernel is 2-dimensional.
0 0 0
Therefore, 1 has geometric multiplicity 2, so this matrix is diagonalizable and similar to D =
 
1
 1 .
3

To write Ain the form SDS −1 ,we


 must
  find
 corresponding eigenvectors of A: the 1-eigenspace

0 2 −2 1 0 −2 2 −2
of A is ker 0 2 −2 = span 0 , 1, and the 3-eigenspace of A is ker  0 0 −2 =
  0 0 0 0 1   0 0 −3 
1 1 0 1 1
span 1. So, one way to diagonalize A is as SDS −1 with S = 0 1 1 and D =  1 .
0 0 1 0 3

4
A matrix An×n is diagonalizable if and only it there is an eigenbasis. In particular, if there are n
distinct eigenvalues, then the matrix is diagonalizable.

5. Recall that A is similar to B if there exists an invertible S such that B = S −1 AS. Show that

(a) similar matrices have the same characteristic polynomial, i.e. the same eigenvalues and algebraic
multiplicities.
Solution. det(B − λI) = det(S −1 AS − λI) = det(S −1 (A − λI)S) = det S det(A − λI) det(S −1 ) =
det(A − λI). Hence they have the same characteristic polynomial.

(b) similar matrices have the same geometric multiplicities.


Solution. Let ~v be an eigenvector of B, and B~v = λ~v . This implies S −1 AS(~v ) = λ~v . So, S~v is
an eigenvector for A because AS(~v ) = λS~v . If we have k linearly independent vectors in Eλ for
B, then we have k linearly independent vectors in Eλ for A.

6. True or false. If the statement is true, explain why; if the statement is false, give a counterexample.

(a) If a 3 × 3 matrix has eigenvalues


 4 withgeometric multiplicity 2 and 5 with geometric multiplicity
4 0 0
1, then it must be similar to 0 4 0.
0 0 5

Solution. True. The fact that A has eigenvalue 4 with geometric multiplicity 2 and 5 with
geometric multiplicity 1 exactly means that we can find three linearly independent eigenvectors,
two with eigenvalue 4 and one with eigenvalue 5.

(b) If an n × n matrix A can be diagonalized as A = SDS −1 with S invertible and D diagonal, then
A3 + 7A2 + 5A + In = S(D3 + 7D2 + 5D + In )S −1 .

Solution. True. We can just write it out:

A3 + 7A2 + 5A + In = AAA + 7AA + 5A + In


= (SDS −1 )(SDS −1 )(SDS −1 ) + 7(SDS −1 )(SDS −1 ) + 5(SDS −1 ) + In
= SD3 S −1 + 7SD2 S −1 + 5SDS −1 + In

Multiplying by a scalar commutes with multiplying by a matrix, so we can rewrite 7SD2 S −1 as


S(7D2 )S −1 ; we can also rewrite In = SIn S −1 :

= SD3 S −1 + S(7D2 )S −1 + S(5D)S −1 + SIn S −1


= S(D3 + 7D2 + 5D + In )S −1

(c) If an invertible matrix A is diagonalizable, then A−1 must be diagonalizable as well.

Solution. True. If A = SDS −1 for a diagonal matrix D, then D−1 is a diagonal matrix (why?)
and A−1 = SD−1 S −1 .

5
   
1 4 2 4 1 6
(d) 0 2 3 and 0 2 0 are similar.
1 2 5 3 2 1

Solution. False. Similar matrices have the same eigenvaules, hence the same traces (sum of the
eigenvaules) and the same determinant (product of eigenvaules). Here, they don’t have the same
trace, so they cannot be similar.
   
4 0 0 4 1 0
(e) 0 4 0 and 0 4 0 are similar.
0 0 5 0 0 5
 
4 0 0
Solution. False. The geometric multiplicity of λ = 4 for 0 4 0 is 2, and the geometric
0 0 5
 
4 1 0
multiplicity of λ = 4 for 0 4 0 is 1. They cannot be similar.
0 0 5

6
• We have two (equivalent) ways of thinking about whether an n × n matrix A is diagonalizable:
saying that A is diagonalizable means that we can write A in the form A = SDS −1 where S is
invertible and D is diagonal, but we often use the equivalent condition that A is diagonalizable
⇐⇒ there is an eigenbasis for A. You should be comfortable with both ways of thinking about
diagonalizable matrices and understand why they are equivalent.
• If a matrix A is diagonalizable, you should be comfortable diagonalizing it; that is, you should
be comfortable finding an invertible matrix S and a diagonal matrix D so that A = SDS −1 .
• You should understand how diagonalizing a matrix A makes it easy to compute powers At of
A. You should also understand how this gives us a new way to think about solving the discrete
dynamical system ~x(t + 1) = A~x(t).
• You should know what it means for two matrices to be similar.

You might also like