Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Matrix Theory PDF

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Matrix Theory, Math6304

Lecture Notes from September 6, 2012


taken by Nathaniel Hammen

Last Time (9/4/12)


Diagonalization: conditions for diagonalization
Eigenvalue Multiplicity: algebraic and geometric multiplicity

1 Further Review
1.1 Warm-up questions
1.1.1 Question. If A ∈ Mn has only one eigenvalue λ (of multiplicity n) and is diagonalizable,
what is A?
Answer. Because A is diagonalizable, there exists an invertible S ∈ Mn such that S −1 AS is
diagonal. In fact, because all eigenvalues are λ, we have
A = S(S −1 AS)S −1 = SλIS −1 = λSS −1 = λI
1.1.2 Question. Let A ∈ Mn and f (t) = det(I + tA). What is f � (t) in terms of A?
Answer. The i, jth entry of (I + tA) is δi,j + tai,j , so
� n

f (t) = det(I + tA) = sgn(σ) (δσ(j),j + taσ(j),j )
σ∈Sn j=1
� n n

� � � �
= sgn(σ) δσ(j),j + t aσ(j),j δσ(i),i + o(t2 )
σ∈Sn j=1 j=1 i�=j

where Sn is the set of permutations of n elements and sgn(σ) is +1 if σ is an even permutation


and -1 if σ is an odd permutation. Differentiating f (t) gives
� n �
� � �
f � (t) = sgn(σ) aσ(j),j δσ(i),i + o(t)
σ∈Sn j=1 i�=j
�n
= aj,j + o(t) = tr(A) + o(t)
j=1

1

This last equality holds because i�=j δσ(i),i is only nonzero when σ(i) = i for all i �= j, which
only occurs when σ is the identity. When we plug 0 into the equation, we get that f � (0) = tr(A).

1.4 Similarity (cont’d)


1.4.27 Theorem. A matrix A ∈ Mn is diagonalizable if and only if the geometric and algebraic
multiplicities of each eigenvalue are equal.

Proof. We begin by noting that if λj and λk are eigenvalues of A with λj �= λk , then their
eigenspaces intersect trivially; that is Eλj ∩ Eλk = {0}. To show this, let x ∈ Eλj ∩ Eλk . Then
λj x = Ax = λk x, which implies x = 0 because λj �= λk . Thus if {v1 , v2 , . . . , vmj } is a basis
for Eλj and {u1 , u2 , . . . , umk } is a basis for Eλk , then {v1 , v2 , . . . , vmj } ∪ {u1 , u2 , . . . , umk } is a
linearly independent set, and forms a basis for Eλj + Eλk . Inductively
� iterating this, we obtain a
basis for Eλ1 + · · · + Eλr , forming a subspace of dimension rj=1 dim(Eλj ). If all algebraic and
geometric multiplicities are equal, then this space has dimension
r
� r

dim(Eλj ) = mj = n.
j=1 j=1

In this case, the subspace equals the entire space, and we have a basis of eigenvectors of A.
Thus, A is diagonalizable.
If not all algebraic and geometric multiplicities are equal, then ∃λk such that dim(Eλk ) < mk .
Assume A is diagonalizable. Then there exists a set S of n linearly independent eigenvectors
of A. For each eigenvalue λj , we know that at most dim(Eλj ) elements in S are eigenvalues
corresponding to λj . Then
r
� r

n = |S| ≤ dim(Eλj ) < mj = n
j=1 j=1

which is a contradiction. Thus, A is not diagonalizable.

1.5 Simultaneous Diagonalization


1.5.28 Definition. Two matrices A, B ∈ Mn are said to be simultaneously diagonalizable if
∃S ∈ Mn such that S is invertible and both S −1 AS and S −1 BS are diagonal matrices.

1.5.29 Remark. If A, B ∈ Mn are simultaneously diagonalizable, then AB = BA.


Proof. Because diagonal matrices commute, we have

AB = S(S −1 AS)(S −1 BS)S −1 = S(S −1 BS)(S −1 AS)S −1 = BA

1.5.30 Question. Is the converse true? That is, if A, B ∈ Mn with A and B diagonalizable and
AB = BA, are A and B simultaneously diagonalizable?
To answer this, we need a lemma.

2
1.5.31
� Lemma.
� Two matrices A ∈ Mn and B ∈ Mm are diagonalizable iff
A 0
C= ∈ Mn+m is diagonalizable.
0 B
Proof. First we assume A ∈ Mn and B ∈ Mm are diagonalizable. Then ∃S1 ∈ Mn , S2 ∈ Mm
such that S1 and S2 are invertible and S −1 AS and S −1 BS are diagonal matrices.
� � � −1 �
S1 0 −1 S1 0
Let S = . Then S =
0 S2 0 S2−1

and � � � �
−1 A 0 S1−1 AS1 0
S S= −1
0 B 0 S2 BS2
which is a diagonal matrix, so C is diagonalizable.
Conversely, assume that C is diagonalizable. Then ∃S ∈ Mn+m such that S is invertible and
D = S −1 CS is a diagonal matrix. Write S = [s1 s2 . . . sn+m ] with each sj ∈ Cn+m . Then,
because SD = CS, each sj is an eigenvector for C. Each sj may then be written as
� �
xj
sj = with xj ∈ Cn and yj ∈ Cm
yj

Then, the block form of C and the fact that Csj = λj sj implies that Axj�= λj�xj and Byj = λj yj .
X
If we let X = [x1 x2 . . . xn+m ] and Y = [y1 y2 . . . yn+m ] then S = . The matrix S is
Y
invertible, so rank(S) = n + m. By the dimensions of X and Y , we also have rank(X) ≤ n
and rank(Y ) ≤ m. When looking at row rank, we see that

n + m = rank(S) ≤ rank(X) + rank(Y ) ≤ n + m

so rank(X) + rank(Y ) = n + m. This can only occur if rank(X) = n and rank(Y ) = m.


Thus X contains n linearly independent columns, each of which is an eigenvector for A, and Y
contains m linearly independent columns, each of which is an eigenvector for B. Thus, we have
bases of eigenvectors of A and B, and we conclude that A and B are diagonalizable.

1.5.32 Theorem. Let A, B ∈ Mn be diagonalizable. Then AB = BA if and only if A and B


are simultaneously diagonalizable.

Proof. We have already shown that if A and B are simultaneously diagonalizable then AB = BA.
All that remains to show is the converse. Assume that AB = BA. Because A is diagonalizable,
∃S ∈ Mn such that S is invertible and D = S −1 AS is diagonal. We may multiply S by an
invertible matrix to permute elements on the diagonal of D, so we may assume without loss of
generality that
 
λ1 I m 1 0 ... 0
 .. 
 0 λ2 I m 2 . 
D= . ..  with λj �= λk ∀j �= k
 .. . 0 
0 ... 0 λr I m r

3
where Im is the m × m identity and mj is the multiplicity of λj . Because AB = BA, we have
(S −1 AS)(S −1 BS) = S −1 ABS = S −1 BAS = (S −1 BS)(S −1 AS).
If we let C = (S −1 BS), then the above gives that DC = CD. If we denote C = [ci,j ]ni,j=1
and D = [di,j ]ni,j=1 , then by the diagonal structure of D we have di,i ci,j = ci,j dj,j . Then
(di,i − dj,j )ci,j = 0 implies that if di,i �= dj,j then ci,j = 0. By the block structure of D, this
implies that C is block diagonal with blocks of the same size. That is
 
C1 0 . . . 0
 .. 
 0 C2 . 
C= . .  with Cj ∈ Mmj ∀j
 .. .. 0 
0 . . . 0 Cr
Because B is diagonalizable, ∃R ∈ Mn such that R is invertible and R−1 BR is diagonal. Then
R−1 SCS −1 R = R−1 SS −1 BSS −1 R = R−1 BR
so C is diagonalizable. From the previous lemma, we deduce that each block Cj is diagonalizable.
Thus, for each j, ∃Tj ∈ Mmj such that each Tj is invertible and Tj−1 Cj Tj is diagonal.
   −1 
T1 0 . . . 0 T1 0 ... 0
 ..   .. 
 0 T2 .  −1  0 T2−1 . 
Let T =  . .  . Then T =  . . 
 .. .. 0   .. .. 0 
0 . . . 0 Tr 0 ... 0 Tr−1
and  
T1−1 C1 T1 0 ... 0
 .. 
−1 −1 −1  0 T2−1 C2 T2 . 
T S BST = T CT =  .. .. 
 . . 0 
0 ... 0 Tr−1 Cr Tr
which is a diagonal matrix. Also
T −1 S −1 AST = T −1 DT
 −1 
T 1 λ1 I m 1 T 1 0 ... 0
 −1 .. 
 0 T 2 λ2 I m 2 T 2 . 
= . .. 
 .
. . 0 
0 ... 0 Tr−1 λr Imr Tr
 
λ1 I m 1 0 ... 0
 .. 
 0 λ2 I m 2 . 
= . .. 
 .. . 0 
0 ... 0 λr I m r
= D.
Thus, T −1 S −1 AST and T −1 S −1 BST are both diagonal matrices, so A and B are simultaneously
diagonalizable by ST ∈ Mn .

4
We want to make this equivalence more general.
1.5.33 Remark. Let A ∈ Mn and consider polynomials

p(A) = p0 I + p1 A + · · · + pr Ar and
q(A) = q0 I + q1 A + · · · + qs As with r, s ∈ N.

Then p(A)q(A) = q(A)p(A). Thus the family of all polynomials of A is commuting. If S ∈ Mn


is invertible, then S −1 Ak S = (S −1 AS)k for any integer value of k, so

S −1 p(A)S = S −1 (p0 I + p1 A + · · · + pr Ar )S
= p0 S −1 IS + p1 S −1 AS + · · · + pr S −1 Ar S
= p0 I + p1 S −1 AS + · · · + pr (S −1 AS)r
= p(S −1 AS)

Thus, if S diagonalizes A, then p(S −1 AS) is a polynomial of a diagonal matrix, which is diagonal.
Thus, if S diagonalizes A, then S −1 p(A)S is diagonal, so S diagonalizes all polynomials of A
simultaneously.

You might also like