Spectral Theorem and Applications

We have seen that diagonalisability is certainly a desirable property for

a matrix to possess. But then, not every matrix is diagonalisable. It is
therefore natural to ask the following:
1 Is there a weaker notion that diagonalisability, which holds true for
every matrix?
2 What are the classes of matrices for which diagonalisability holds
true?
We shall see that both the questions have nice answers for square
matrices over C. Moreover, we can in fact, replace diagonalisability of
a matrix by the following stronger notion that is more natural in the
context of matrices over C. Recall that a n × n matrix C is said to be
unitary if C ∗ = C −1 , i.e., if CC ∗ = I.

Definition
A n × n matrix A with entries in C is said to be unitarily diagonalisable if
there exists a unitary n × n matrix C such that C ∗ AC is a diagonal
matrix.
1/33
Congruence of Matrices
Just as a matrix being diagonalisable is the same thing as saying it is
similar to a diagonal matrix, a matrix being unitarily diagonalisable is
the same thing as saying it is “congruent” to a diagonal matrix, where
the notion of congruence in the class of square matrices with entries in
C is defined as follows.
Definition
Two n × n matrices A and B are said to be congruent if there exists a
unitary n × n matrix C such that C ∗ AC = B.

In case, the matrices A, B, C above have real entries, then to say that C is
unitary is the same as saying that C is (real) orthogonal and the relation
C ∗ AC = B can be written as C T AC = B.
Now the first question above admits a nice answer by using the following
notion that is weaker than the notion of diagonalisability.

2/33
Congruence of Matrices
Just as a matrix being diagonalisable is the same thing as saying it is
similar to a diagonal matrix, a matrix being unitarily diagonalisable is
the same thing as saying it is “congruent” to a diagonal matrix, where
the notion of congruence in the class of square matrices with entries in
C is defined as follows.
Definition
Two n × n matrices A and B are said to be congruent if there exists a
unitary n × n matrix C such that C ∗ AC = B.

In case, the matrices A, B, C above have real entries, then to say that C is
unitary is the same as saying that C is (real) orthogonal and the relation
C ∗ AC = B can be written as C T AC = B.
Now the first question above admits a nice answer by using the following
notion that is weaker than the notion of diagonalisability.
Definition
We say A is triangularizable if there exists an invertible matrix C such
that C −1 AC is upper triangular.
2/33
Congruence of Matrices
Just as a matrix being diagonalisable is the same thing as saying it is
similar to a diagonal matrix, a matrix being unitarily diagonalisable is
the same thing as saying it is “congruent” to a diagonal matrix, where
the notion of congruence in the class of square matrices with entries in
C is defined as follows.
Definition
Two n × n matrices A and B are said to be congruent if there exists a
unitary n × n matrix C such that C ∗ AC = B.

In case, the matrices A, B, C above have real entries, then to say that C is
unitary is the same as saying that C is (real) orthogonal and the relation
C ∗ AC = B can be written as C T AC = B.
Now the first question above admits a nice answer by using the following
notion that is weaker than the notion of diagonalisability.
Definition
We say A is triangularizable if there exists an invertible matrix C such
that C −1 AC is upper triangular.
2/33
Triangularization

Here is the promised nice answer to Q. 1 and this may be thought of

as “poor man’s version of diagonalisation".

Proposition
Over the complex numbers, every square matrix is congruent to an
upper triangular matrix.

3/33
Triangularization

Here is the promised nice answer to Q. 1 and this may be thought of

as “poor man’s version of diagonalisation".

Proposition
Over the complex numbers, every square matrix is congruent to an
upper triangular matrix.

Proof: Let A be a n × n matrix with complex entries. We have to find a

unitary matrix C such that C ∗ AC is upper triangular.

3/33
Triangularization

Here is the promised nice answer to Q. 1 and this may be thought of

as “poor man’s version of diagonalisation".

Proposition
Over the complex numbers, every square matrix is congruent to an
upper triangular matrix.

Proof: Let A be a n × n matrix with complex entries. We have to find a

unitary matrix C such that C ∗ AC is upper triangular. We shall prove this by
induction. For n = 1 there is nothing to prove. Assume the result for n − 1.

3/33
Triangularization

Here is the promised nice answer to Q. 1 and this may be thought of

as “poor man’s version of diagonalisation".

Proposition
Over the complex numbers, every square matrix is congruent to an
upper triangular matrix.

Proof: Let A be a n × n matrix with complex entries. We have to find a

unitary matrix C such that C ∗ AC is upper triangular. We shall prove this by
induction. For n = 1 there is nothing to prove. Assume the result for n − 1.
Let µ be an eigenvalue of A and let v ∈ Cn be such that Av1 = µv1 . (Here
we need to work with complex numbers; real numbers won’t do. why?) We
can choose v1 to be of norm 1. We can then complete it to an orthonormal
basis {v1 , . . . , vn }. (Here we use Gram-Schmidt.)

3/33
Triangularization

Here is the promised nice answer to Q. 1 and this may be thought of

as “poor man’s version of diagonalisation".

Proposition
Over the complex numbers, every square matrix is congruent to an
upper triangular matrix.

Proof: Let A be a n × n matrix with complex entries. We have to find a

unitary matrix C such that C ∗ AC is upper triangular. We shall prove this by
induction. For n = 1 there is nothing to prove. Assume the result for n − 1.
Let µ be an eigenvalue of A and let v ∈ Cn be such that Av1 = µv1 . (Here
we need to work with complex numbers; real numbers won’t do. why?) We
can choose v1 to be of norm 1. We can then complete it to an orthonormal
basis {v1 , . . . , vn }. (Here we use Gram-Schmidt.) We then take C1 to be the
matrix whose i th column is vi . Then as seen earlier, C1 is a unitary matrix.

3/33
Put A1 = C1−1 AC1 . Then

A1 e1 = C1−1 AC1 e1 = C1−1 Av1 = µC1−1 (v1 ) = µe1 .

This shows that the first column of A1 has all entries zero except the
first one which is equal to µ.

4/33
Put A1 = C1−1 AC1 . Then

A1 e1 = C1−1 AC1 e1 = C1−1 Av1 = µC1−1 (v1 ) = µe1 .

This shows that the first column of A1 has all entries zero except the
first one which is equal to µ. Let B be the matrix obtained from A1 by
cutting down the first row and the first column, so that A1 is a block
matrix of the form  
µ ...
A1 =
0n−1 B
where 0n−1 is the column of zeros of size n − 1. By induction, there
exists a (n − 1) × (n − 1) unitary matrix M such that −1
 t
 M BM is an
1 0n−1
upper triangular matrix. Put M1 = .
0n−1 M

4/33
Put A1 = C1−1 AC1 . Then

A1 e1 = C1−1 AC1 e1 = C1−1 Av1 = µC1−1 (v1 ) = µe1 .

This shows that the first column of A1 has all entries zero except the
first one which is equal to µ. Let B be the matrix obtained from A1 by
cutting down the first row and the first column, so that A1 is a block
matrix of the form  
µ ...
A1 =
0n−1 B
where 0n−1 is the column of zeros of size n − 1. By induction, there
exists a (n − 1) × (n − 1) unitary matrix M such that −1
 t
 M BM is an
1 0n−1
upper triangular matrix. Put M1 = .
0n−1 M
Then M1 is unitary and hence C = C1 M1 is also unitary.

4/33
Put A1 = C1−1 AC1 . Then

A1 e1 = C1−1 AC1 e1 = C1−1 Av1 = µC1−1 (v1 ) = µe1 .

This shows that the first column of A1 has all entries zero except the
first one which is equal to µ. Let B be the matrix obtained from A1 by
cutting down the first row and the first column, so that A1 is a block
matrix of the form  
µ ...
A1 =
0n−1 B
where 0n−1 is the column of zeros of size n − 1. By induction, there
exists a (n − 1) × (n − 1) unitary matrix M such that −1
 t
 M BM is an
1 0n−1
upper triangular matrix. Put M1 = .
0n−1 M
Then M1 is unitary and hence C = C1 M1 is also unitary. Clearly
C −1 AC = M1−1 C1−1 AC1 M1 = M1−1 A1 M1 which is of the form
0tn−1 0tn−1
   
1 µ ... 1
0n−1 M −1 0n−1 B 0n−1 M
and hence is upper triangular.
4/33
Remark
Assume now that A is a real matrix with all its eigenvalues real. Then it
follows that we can choose the eigenvector v1 to be a real vector and
then complete this into a basis for Rn .

5/33
Remark
Assume now that A is a real matrix with all its eigenvalues real. Then it
follows that we can choose the eigenvector v1 to be a real vector and
then complete this into a basis for Rn . Thus the matrix C1
corresponding to this basis will have real entries.

5/33
Remark
Assume now that A is a real matrix with all its eigenvalues real. Then it
follows that we can choose the eigenvector v1 to be a real vector and
then complete this into a basis for Rn . Thus the matrix C1
corresponding to this basis will have real entries. By induction M will
have real entries and hence the product C = MC1 will also have real
entries. Thus we have proved:

5/33
Remark
Assume now that A is a real matrix with all its eigenvalues real. Then it
follows that we can choose the eigenvector v1 to be a real vector and
then complete this into a basis for Rn . Thus the matrix C1
corresponding to this basis will have real entries. By induction M will
have real entries and hence the product C = MC1 will also have real
entries. Thus we have proved:

Proposition
For a real square matrix A with all its eigenvalues real, there exists an
orthogonal matrix C such that C t AC is upper triangular.

5/33
Normal Matrices
Definition
A square matrix A is called normal if A∗ A = AA∗ .

6/33
Normal Matrices
Definition
A square matrix A is called normal if A∗ A = AA∗ .

Remark
(i) Normality is congruence invariant. This means that if C is unitary
and A is normal then C −1 AC is also normal. This is easy to verify.

6/33
Normal Matrices
Definition
A square matrix A is called normal if A∗ A = AA∗ .

Remark
(i) Normality is congruence invariant. This means that if C is unitary
and A is normal then C −1 AC is also normal. This is easy to verify.
(ii) Any diagonal matrix is normal. Therefore it follows that normality is
necessary for diagonalization. Amazingly, it turns out to be sufficient.
That is the reason to define this concept.

6/33
Normal Matrices
Definition
A square matrix A is called normal if A∗ A = AA∗ .

Remark
(i) Normality is congruence invariant. This means that if C is unitary
and A is normal then C −1 AC is also normal. This is easy to verify.
(ii) Any diagonal matrix is normal. Therefore it follows that normality is
necessary for diagonalization. Amazingly, it turns out to be sufficient.
That is the reason to define this concept.
(iii) Observe that product
 of two normal matrices
 may not be normal.
0 1 1 0
For example take A = ; B= .
1 0 0 2

6/33
Normal Matrices
Definition
A square matrix A is called normal if A∗ A = AA∗ .

Remark
(i) Normality is congruence invariant. This means that if C is unitary
and A is normal then C −1 AC is also normal. This is easy to verify.
(ii) Any diagonal matrix is normal. Therefore it follows that normality is
necessary for diagonalization. Amazingly, it turns out to be sufficient.
That is the reason to define this concept.
(iii) Observe that product
 of two normal matrices
 may not be normal.
0 1 1 0
For example take A = ; B= .
1 0 0 2
(iv) Certainly Hermitian matrices are normal. Of course there are
normal  matrices
 which are not Hermitian. For example take
ı 0
A= .
0 −ı
6/33
Lemma
For a normal matrix A we have kAxk2 = kA∗ xk2 for all x ∈ Cn .

7/33
Lemma
For a normal matrix A we have kAxk2 = kA∗ xk2 for all x ∈ Cn .

Proof: hAx, Axi = (Ax)∗ (Ax) = x∗ A∗ Ax = x∗ AA∗ x = (A∗ x)∗ (A∗ x) =

hA∗ x, A∗ xi = kA∗ xk2 . ♠

7/33
Lemma
For a normal matrix A we have kAxk2 = kA∗ xk2 for all x ∈ Cn .

Proof: hAx, Axi = (Ax)∗ (Ax) = x∗ A∗ Ax = x∗ AA∗ x = (A∗ x)∗ (A∗ x) =

hA∗ x, A∗ xi = kA∗ xk2 . ♠

Lemma
If A is normal, then v is an eigenvector of A with eigenvalue µ iff v is an
eigenvector of A∗ with eigenvalue µ.

7/33
Lemma
For a normal matrix A we have kAxk2 = kA∗ xk2 for all x ∈ Cn .

Proof: hAx, Axi = (Ax)∗ (Ax) = x∗ A∗ Ax = x∗ AA∗ x = (A∗ x)∗ (A∗ x) =

hA∗ x, A∗ xi = kA∗ xk2 . ♠

Lemma
If A is normal, then v is an eigenvector of A with eigenvalue µ iff v is an
eigenvector of A∗ with eigenvalue µ.

Proof: Observe that if A is normal then A − µI is also normal. Now

(A − µI)(v) = 0 iff k(A − µI)(v)k = 0 iff k(A − µI)∗ vk = 0 iff
(A∗ − µI)(v) = 0. ♠

7/33
Proposition
An upper triangular normal matrix is diagonal.

8/33
Proposition
An upper triangular normal matrix is diagonal.

Proof: Let A be an upper triangular normal matrix. Inductively we shall

show that aij = 0 for j > i.

8/33
Proposition
An upper triangular normal matrix is diagonal.

Proof: Let A be an upper triangular normal matrix. Inductively we shall

show that aij = 0 for j > i. We have Ae1 = a11 e1 . Hence kAe1 k2 = |a11 |2 .

8/33
Proposition
An upper triangular normal matrix is diagonal.

Proof: Let A be an upper triangular normal matrix. Inductively we shall

show that aij = 0 for j > i. We have Ae1 = a11 e1 . Hence kAe1 k2 = |a11 |2 .
On the other hand this is equal to kA∗ e1 k = |a11 |2 + |a12 |2 + · · · + |a1n |2 .

8/33
Proposition
An upper triangular normal matrix is diagonal.

Proof: Let A be an upper triangular normal matrix. Inductively we shall

show that aij = 0 for j > i. We have Ae1 = a11 e1 . Hence kAe1 k2 = |a11 |2 .
On the other hand this is equal to kA∗ e1 k = |a11 |2 + |a12 |2 + · · · + |a1n |2 .
Hence a12 = a13 = · · · = a1n = 0. Inductively suppose we have shown
aij = 0 for j > i for all 1 ≤ i ≤ k − 1. Then it follows that Aek = akk ek .

8/33
Proposition
An upper triangular normal matrix is diagonal.

Proof: Let A be an upper triangular normal matrix. Inductively we shall

show that aij = 0 for j > i. We have Ae1 = a11 e1 . Hence kAe1 k2 = |a11 |2 .
On the other hand this is equal to kA∗ e1 k = |a11 |2 + |a12 |2 + · · · + |a1n |2 .
Hence a12 = a13 = · · · = a1n = 0. Inductively suppose we have shown
aij = 0 for j > i for all 1 ≤ i ≤ k − 1. Then it follows that Aek = akk ek .
Exactly as in the first case, this implies that
kA∗ ek k = |ak ,k |2 + |ak ,k +1 |2 + · · · + |ak ,n |2 = |ak ,k |2 . Hence
ak ,k +1 = · · · = ak ,n = 0. ♠

8/33
Here is an answer to the second question raised at the beginning of
this topic. This result, called the Spectral Theorem, is one of the most
remarkable results in linear algebra, and it says that every normal
matrix can be unitarily diagonalised. An important corollary is that a
real symmetric matrix is orthogonally diagonalisable.
Theorem (Spectral Theorem)
Given any normal matrix A, there exists a unitary matrix C such that
C ∗ AC is a diagonal matrix.

9/33
Here is an answer to the second question raised at the beginning of
this topic. This result, called the Spectral Theorem, is one of the most
remarkable results in linear algebra, and it says that every normal
matrix can be unitarily diagonalised. An important corollary is that a
real symmetric matrix is orthogonally diagonalisable.
Theorem (Spectral Theorem)
Given any normal matrix A, there exists a unitary matrix C such that
C ∗ AC is a diagonal matrix.

Corollary
Every Hermitian matrix A is congruent to a diagonal matrix. A real
symmetric matrix is real-congruent to a diagonal matrix.

9/33
Here is an answer to the second question raised at the beginning of
this topic. This result, called the Spectral Theorem, is one of the most
remarkable results in linear algebra, and it says that every normal
matrix can be unitarily diagonalised. An important corollary is that a
real symmetric matrix is orthogonally diagonalisable.
Theorem (Spectral Theorem)
Given any normal matrix A, there exists a unitary matrix C such that
C ∗ AC is a diagonal matrix.

Corollary
Every Hermitian matrix A is congruent to a diagonal matrix. A real
symmetric matrix is real-congruent to a diagonal matrix.

Proof of Corollary: (a) We observe that a Hermitian matrix is normal and

apply the above theorem.

9/33
Here is an answer to the second question raised at the beginning of
this topic. This result, called the Spectral Theorem, is one of the most
remarkable results in linear algebra, and it says that every normal
matrix can be unitarily diagonalised. An important corollary is that a
real symmetric matrix is orthogonally diagonalisable.
Theorem (Spectral Theorem)
Given any normal matrix A, there exists a unitary matrix C such that
C ∗ AC is a diagonal matrix.

Corollary
Every Hermitian matrix A is congruent to a diagonal matrix. A real
symmetric matrix is real-congruent to a diagonal matrix.

Proof of Corollary: (a) We observe that a Hermitian matrix is normal and

apply the above theorem.
(b) We first recall that for a real symmetric matrix, all eigenvalues are real.
Moreover, eignevectors can be chosen to be real as well. Hence we can
choose C to be an orthogonal matrix. Thus the Spectral Theorem gives the
desired result.
9/33
Quadratic forms and their diagonalization

Definition
Let A = (aij ) be an n × n real matrix . The function Q : Rn → R defined
by :
Xn X n
Q(x) = aij xi xj , X = (x1 , x2 , . . . , xn )t ∈ Rn
i=1 j=1

is called the quadratic form associated with A.

10/33
Quadratic forms and their diagonalization

Definition
Let A = (aij ) be an n × n real matrix . The function Q : Rn → R defined
by :
Xn X n
Q(x) = aij xi xj , X = (x1 , x2 , . . . , xn )t ∈ Rn
i=1 j=1

is called the quadratic form associated with A. If

A = diag (λ1 , λ2 , . . . , λn ) then Q(X ) = λ1 x12 + λ2 x22 + · · · + λn xn2 is
called a diagonal form.

10/33
Proposition
 
x1
 x2 
Q(X ) = [x1 , x2 , . . . , xn ]A   = X t AX where X = (x1 , x2 , . . . , xn )t .
 
..
 . 
xn

11/33
Proposition
 
x1
 x2 
Q(X ) = [x1 , x2 , . . . , xn ]A   = X t AX where X = (x1 , x2 , . . . , xn )t .
 
..
 . 
xn

Proof:
 n
X


x1

 a1j xj 
 j=1 
 x2  
..

[x1 , x2 , . . . , xn ]A  = [x , x , . . . , x ]
   
..
 n .
 1 2 n  
 .
 X


xn  anj xj 
j=1

11/33
Proposition
 
x1
 x2 
Q(X ) = [x1 , x2 , . . . , xn ]A   = X t AX where X = (x1 , x2 , . . . , xn )t .
 
..
 . 
xn

Proof:
 n
X


x1

 a1j xj 
 j=1 
 x2  
..

[x1 , x2 , . . . , xn ]A  = [x , x , . . . , x ]
   
..
 n .
 1 2 n  
 .
 X


xn  anj xj 
j=1
n
X n
X
= a1j xj x1 + · · · + anj xj xn
j=1 i=1
n X
X n
= aij xi xj
11/33
Example
   
1 1 x
(1) A = , X = .
3 5 y

12/33
Example
   
1 1 x
(1) A = , X = . Then
3 5 y
    
t 1 1 x x +y
X AX = [x, y ] = [x, y ] = x12 + 4xy + 5y 2 .
3 5 y 3x + 5y

12/33
Example
   
1 1 x
(1) A = , X = . Then
3 5 y
    
t 1 1 x x +y
X AX = [x, y ] = [x, y ] = x12 + 4xy + 5y 2 .
3 5 y 3x + 5y
   
1 2 x
(2) Let B = , X = .
2 5 y

12/33
Example
   
1 1 x
(1) A = , X = . Then
3 5 y
    
t 1 1 x x +y
X AX = [x, y ] = [x, y ] = x12 + 4xy + 5y 2 .
3 5 y 3x + 5y
   
1 2 x
(2) Let B = , X = . Then
2 5 y
    
t 1 2 x x + 2y
X BX = [x, y ] = [x, y ] = x12 + 4xy + 5x22 .
2 5 y 2x + 5y

12/33
Example
   
1 1 x
(1) A = , X = . Then
3 5 y
    
t 1 1 x x +y
X AX = [x, y ] = [x, y ] = x12 + 4xy + 5y 2 .
3 5 y 3x + 5y
   
1 2 x
(2) Let B = , X = . Then
2 5 y
    
t 1 2 x x + 2y
X BX = [x, y ] = [x, y ] = x12 + 4xy + 5x22 .
2 5 y 2x + 5y

Notice that A and B give rise to same Q(x) and B = 12 (A + At ) is a

symmetric matrix.

12/33
Proposition
For any n × n matrix A and the column vector X = (x1 , x2 , . . . , xn )t ,

1
X t AX = X t BX where B = (A + At ).
2
Hence every quadratic form is associated with a symmetric matrix.

13/33
Proposition
For any n × n matrix A and the column vector X = (x1 , x2 , . . . , xn )t ,

1
X t AX = X t BX where B = (A + At ).
2
Hence every quadratic form is associated with a symmetric matrix.

Proof: X t AX is a 1 × 1 matrix. Hence X t AX = X t At X = (X t AX )t . Hence

1 t 1 1
X t AX = X AX + X t At X = X t A + At X = X t BX .


2 2 2
♠

13/33
Quadratic forms and their diagonalization
We now show how the spectral theorem helps us in converting a
quadratic form into a diagonal form.

14/33
Quadratic forms and their diagonalization
We now show how the spectral theorem helps us in converting a
quadratic form into a diagonal form.

Theorem
Let X t AX be a quadratic form associated with a real symmetric matrix
A. Let U be an orthogonal matrix such that
U t AU = diag (λ1 , λ2 , . . . , λn ).

14/33
Quadratic forms and their diagonalization
We now show how the spectral theorem helps us in converting a
quadratic form into a diagonal form.

Theorem
Let X t AX be a quadratic form associated with a real symmetric matrix
A. Let U be an orthogonal matrix such that
U t AU = diag (λ1 , λ2 , . . . , λn ). Then

X t AX = λ1 y12 + λ2 y22 + · · · + λn yn2 ,

where    
x1 y1
 x2   y2 
X = =U  = UY .
   
.. ..
 .   . 
xn yn

14/33
Proof: Since X = UY ,

X t AX = (UY )t A(UY ) = Y t (U t AU)Y .

15/33
Proof: Since X = UY ,

X t AX = (UY )t A(UY ) = Y t (U t AU)Y .

Since U t AU = diag (λ1 , λ2 , . . . , λn ), we get

   
λ1 y1
 λ2   y2 
X t AX = [y1 , y2 , . . . , yn ] 
   
..   .. 
 .   . 
λn yn

15/33
Proof: Since X = UY ,

X t AX = (UY )t A(UY ) = Y t (U t AU)Y .

Since U t AU = diag (λ1 , λ2 , . . . , λn ), we get

   
λ1 y1
 λ2   y2 
X t AX = [y1 , y2 , . . . , yn ] 
   
..   .. 
 .   . 
λn yn
= λ1 y12 + λ2 y22 + · · · + λn yn2 .

15/33
Example
Let us determine the orthogonal matrix U which reduces the quadratic
form Q(X ) = 2x 2 + 4xy + 5y 2 to a diagonal form.

16/33
Example
Let us determine the orthogonal matrix U which reduces the quadratic
form Q(X ) = 2x 2 + 4xy + 5y 2 to a diagonal form. We write
     
2 2 x t x
Q(X ) = [x, y ] = X AX with X = .
2 5 y y

16/33
Example
Let us determine the orthogonal matrix U which reduces the quadratic
form Q(X ) = 2x 2 + 4xy + 5y 2 to a diagonal form. We write
     
2 2 x t x
Q(X ) = [x, y ] = X AX with X = .
2 5 y y

The symmetric matrix A can be diagonalized. The eigenvalues of A are

λ1 = 1 and λ2 = 6. An orthonormal set of eigenvectors for λ1 and λ2 is
   
1 2 1 1
v1 = √ and v2 = √ .
5 −1 5 2

16/33
Example
Let us determine the orthogonal matrix U which reduces the quadratic
form Q(X ) = 2x 2 + 4xy + 5y 2 to a diagonal form. We write
     
2 2 x t x
Q(X ) = [x, y ] = X AX with X = .
2 5 y y

The symmetric matrix A can be diagonalized. The eigenvalues of A are

λ1 = 1 and λ2 = 6. An orthonormal set of eigenvectors for λ1 and λ2 is
   
1 2 1 1
v1 = √ and v2 = √ .
5 −1 5 2
 
√1
2 1
Hence U = . The change of variables equations are
5 −1 2
   
x u
=U .
y v

16/33
Example
Let us determine the orthogonal matrix U which reduces the quadratic
form Q(X ) = 2x 2 + 4xy + 5y 2 to a diagonal form. We write
     
2 2 x t x
Q(X ) = [x, y ] = X AX with X = .
2 5 y y

The symmetric matrix A can be diagonalized. The eigenvalues of A are

λ1 = 1 and λ2 = 6. An orthonormal set of eigenvectors for λ1 and λ2 is
   
1 2 1 1
v1 = √ and v2 = √ .
5 −1 5 2
 
√1
2 1
Hence U = . The change of variables equations are
5 −1 2
   
x u
=U . The diagonal form is:
y v
 
1 0
[u, v ] [u, v ]T = u 2 + 6v 2 . Check that U t AU = diag (1, 6).
0 6
16/33
Conic Sections and quadric surfaces
A conic section is the locus in the Cartesian plane R2 of an equation of
the form

ax 2 + bxy + cy 2 + dx + ey + f = 0. (1)

17/33
Conic Sections and quadric surfaces
A conic section is the locus in the Cartesian plane R2 of an equation of
the form

ax 2 + bxy + cy 2 + dx + ey + f = 0. (1)

It can be proved that this equation represents one of the following: (i)
the empty set
(ii) a single point
(iii) one or two straight lines
(iv) an ellipse
(v) an hyperbola and
(vi) a parabola.

17/33
Conic Sections and quadric surfaces
A conic section is the locus in the Cartesian plane R2 of an equation of
the form

ax 2 + bxy + cy 2 + dx + ey + f = 0. (1)

It can be proved that this equation represents one of the following: (i)
the empty set
(ii) a single point
(iii) one or two straight lines
(iv) an ellipse
(v) an hyperbola and
(vi) a parabola.

The second degree part of (1)

Q(x, y ) = ax 2 + bxy + cy 2

is a quadratic form. This determines the type of the conic.

17/33
We can write the equation (1) into matrix form after setting
x = x, y = y :
     
a b/2 x x
[x, y ] + [d, e] +f = 0 (2)
b/2 c y y
 
a b/2
Write A = . Let U = [v1 , v2 ] be an orthogonal matrix
b/2 c
whose column vectors v1 and v2 are eigenvectors of A with
eigenvalues λ1 and λ2 . Apply the change of variables
   
x u
X = =U
y v
to diagonalize the quadratic form Q(x, y ) to the diagonal form
λ1 y12 + λ2 y22 .

18/33
We can write the equation (1) into matrix form after setting
x = x, y = y :
     
a b/2 x x
[x, y ] + [d, e] +f = 0 (2)
b/2 c y y
 
a b/2
Write A = . Let U = [v1 , v2 ] be an orthogonal matrix
b/2 c
whose column vectors v1 and v2 are eigenvectors of A with
eigenvalues λ1 and λ2 . Apply the change of variables
   
x u
X = =U
y v
to diagonalize the quadratic form Q(x, y ) to the diagonal form
λ1 y12 + λ2 y22 . The orthonormal basis {v1 , v2 } determines a new set of
coordinate axes with respect to which the locus of the equation
[x, y ]A[x, y ]T + B[x, y ]T + f = 0 with B = [d, e] is same as the locus of
the equation
0 = [u, v ] diag (λ1 , λ2 )[u, v ]T + (BU)[u, v ]T + f

18/33
We can write the equation (1) into matrix form after setting
x = x, y = y :
     
a b/2 x x
[x, y ] + [d, e] +f = 0 (2)
b/2 c y y
 
a b/2
Write A = . Let U = [v1 , v2 ] be an orthogonal matrix
b/2 c
whose column vectors v1 and v2 are eigenvectors of A with
eigenvalues λ1 and λ2 . Apply the change of variables
   
x u
X = =U
y v
to diagonalize the quadratic form Q(x, y ) to the diagonal form
λ1 y12 + λ2 y22 . The orthonormal basis {v1 , v2 } determines a new set of
coordinate axes with respect to which the locus of the equation
[x, y ]A[x, y ]T + B[x, y ]T + f = 0 with B = [d, e] is same as the locus of
the equation
0 = [u, v ] diag (λ1 , λ2 )[u, v ]T + (BU)[u, v ]T + f
= λ1 u 2 + λ2 v 2 + [d, e][v1 , v2 ][u, v ]T + f . (3)
18/33
If the conic determined by (3) is not degenerate i.e., not an empty set,
a point, nor line(s) then signs of λ1 and λ2 determine whether it is a
parabola, an hyperbola or an ellipse.

19/33
If the conic determined by (3) is not degenerate i.e., not an empty set,
a point, nor line(s) then signs of λ1 and λ2 determine whether it is a
parabola, an hyperbola or an ellipse. The equation (1) will represent

(1) ellipse if λ1 λ2 > 0

(2) hyperbola if λ1 λ2 < 0

(3) parabola if λ1 λ2 = 0

19/33
Example
(1) 2x 2 + 4xy + 5y 2 + 4x + 13y − 1/4 = 0.
We have earlier diagonalized the quadratic form 2x 2 + 4xy + 5y 2 .

20/33
Example
(1) 2x 2 + 4xy + 5y 2 + 4x + 13y − 1/4 = 0.
We have earlier diagonalized the quadratic form 2x 2 + 4xy + 5y 2 . The
associated symmetric matrix, the eigenvectors and eigenvalues are
displayed in the equation of diagonalization :
     
t 1 2 −1 2 2 1 2 1
U AU = √ √
5 1 2 2 5 5 −1 2
 
1 0
= .
0 6

20/33
Example
(1) 2x 2 + 4xy + 5y 2 + 4x + 13y − 1/4 = 0.
We have earlier diagonalized the quadratic form 2x 2 + 4xy + 5y 2 . The
associated symmetric matrix, the eigenvectors and eigenvalues are
displayed in the equation of diagonalization :
     
t 1 2 −1 2 2 1 2 1
U AU = √ √
5 1 2 2 5 5 −1 2
 
1 0
= .
0 6
√
Set t = 1/ 5 for convenience. Then the change of coordinates
equations are :      
x 2t t u
= ,
y −t 2t v

20/33
Example
i.e., x = t(2u + v ) and y = t(−u + 2v ). Substitute these into the
original equation to get
√ √ 1
u 2 + 6v 2 − 5u + 6 5v − = 0.
4

21/33
Example
i.e., x = t(2u + v ) and y = t(−u + 2v ). Substitute these into the
original equation to get
√ √ 1
u 2 + 6v 2 − 5u + 6 5v − = 0.
4
Complete the square to write this as

1√ 2 1√ 2
(u − 5) + 6(v + 5) = 9.
2 2

21/33
Example
i.e., x = t(2u + v ) and y = t(−u + 2v ). Substitute these into the
original equation to get
√ √ 1
u 2 + 6v 2 − 5u + 6 5v − = 0.
4
Complete the square to write this as

1√ 2 1√ 2
(u − 5) + 6(v + 5) = 9.
2 2
√ √
This is an equation of ellipse with center ( 12 5, − 12 5) in the uv -plane.

21/33
Example
The u-axis and v -axis are determined by the eigenvectors v1 and v2
as indicated in the following figure :

22/33
Example
2 2
 − 4xy − y − 4x + 10y − 13 = 0. Here, the matrix
(2) 2x
2 −2
A= gives the quadratic part of the equation.
−2 −1

23/33
Example
2 2
 − 4xy − y − 4x + 10y − 13 = 0. Here, the matrix
(2) 2x
2 −2
A= gives the quadratic part of the equation. We write the
−2 −1
equation in matrix form as
    
2 −2 x x
[x, y ] + [−4, 10] − 13 = 0.
−2 −1 y y

23/33
Example
2 2
 − 4xy − y − 4x + 10y − 13 = 0. Here, the matrix
(2) 2x
2 −2
A= gives the quadratic part of the equation. We write the
−2 −1
equation in matrix form as
    
2 −2 x x
[x, y ] + [−4, 10] − 13 = 0.
−2 −1 y y

√
Let t = 1/ 5. The eigenvalues of A are λ1 = 3, λ2 = −2. An
orthonormal set of eigenvectors is v1 = t(2, −1)t and v2 = t(1, 2)t . Put
     
2 1 x u
U=t and =U .
−1 2 y v

23/33
Example
2 2
 − 4xy − y − 4x + 10y − 13 = 0. Here, the matrix
(2) 2x
2 −2
A= gives the quadratic part of the equation. We write the
−2 −1
equation in matrix form as
    
2 −2 x x
[x, y ] + [−4, 10] − 13 = 0.
−2 −1 y y

√
Let t = 1/ 5. The eigenvalues of A are λ1 = 3, λ2 = −2. An
orthonormal set of eigenvectors is v1 = t(2, −1)t and v2 = t(1, 2)t . Put
     
2 1 x u
U=t and =U .
−1 2 y v

The transformed equation becomes

3u 2 − 2v 2 − 4t(2u + v ) + 10t(−u + 2v ) − 13 = 0

23/33
Example
or
3u 2 − 2v 2 − 18tu + 16tv − 13 = 0.

24/33
Example
or
3u 2 − 2v 2 − 18tu + 16tv − 13 = 0.
Complete the square in u and v to get

3(u − 3t)2 − 2(v − 4t)2 = 12

or
(u − 3t)2 (v − 4t)2
− = 1.
4 6

24/33
Example
or
3u 2 − 2v 2 − 18tu + 16tv − 13 = 0.
Complete the square in u and v to get

3(u − 3t)2 − 2(v − 4t)2 = 12

or
(u − 3t)2 (v − 4t)2
− = 1.
4 6
This represents a hyperbola with center (3t, 4t) in the uv -plane.

24/33
Example
or
3u 2 − 2v 2 − 18tu + 16tv − 13 = 0.
Complete the square in u and v to get

3(u − 3t)2 − 2(v − 4t)2 = 12

or
(u − 3t)2 (v − 4t)2
− = 1.
4 6
This represents a hyperbola with center (3t, 4t) in the uv -plane. The
eigenvectors v1 and v2 determine the directions of positive u and v
axes.

24/33
Example

25/33
Example
(3) 9x 2 + 24xy + 16xy 2 − 20x + 15y = 0  
9 12
The symmetric matrix for the quadratic part is A = .
12 16

26/33
Example
(3) 9x 2 + 24xy + 16xy 2 − 20x + 15y = 0  
9 12
The symmetric matrix for the quadratic part is A = . The
12 16
eigenvalues are λ1 = 25, λ2 = 0. An orthonormal set of eigenvectors is
v1 = a(3, 4)t , v2 = a(−4, 3)t where a = 1/5.

26/33
Example
(3) 9x 2 + 24xy + 16xy 2 − 20x + 15y = 0  
9 12
The symmetric matrix for the quadratic part is A = . The
12 16
eigenvalues are λ1 = 25, λ2 = 0. An orthonormal set of eigenvectors is
v1 = a(3, 4)t , v2 = a(−4, 3)t where
 a =1/5. An orthogonal
3 −4
diagonalizing matrix is U = a .
4 3

26/33
Example
(3) 9x 2 + 24xy + 16xy 2 − 20x + 15y = 0  
9 12
The symmetric matrix for the quadratic part is A = . The
12 16
eigenvalues are λ1 = 25, λ2 = 0. An orthonormal set of eigenvectors is
v1 = a(3, 4)t , v2 = a(−4, 3)t where
 a =1/5. An orthogonal
3 −4
diagonalizing matrix is U = a . The equations of change of
4 3
coordinates are
   
x u
=U i.e., x = a(3u − 4v ), y = a(4u + 3v ).
y v

26/33
Example
(3) 9x 2 + 24xy + 16xy 2 − 20x + 15y = 0  
9 12
The symmetric matrix for the quadratic part is A = . The
12 16
eigenvalues are λ1 = 25, λ2 = 0. An orthonormal set of eigenvectors is
v1 = a(3, 4)t , v2 = a(−4, 3)t where
 a =1/5. An orthogonal
3 −4
diagonalizing matrix is U = a . The equations of change of
4 3
coordinates are
   
x u
=U i.e., x = a(3u − 4v ), y = a(4u + 3v ).
y v

The equation in uv -plane is u 2 + v = 0. This is an equation of parabola

with its vertex at the origin.

26/33
Quadric Surfaces

Let A be a 3 × 3 real symmetric matrix. The locus of the equation

   
x x
[x, y , z]A  y  + [b1 , b2 , b3 ]  y  + c = 0 (4)
z z

in three variables is called a quadric surface.

27/33
Quadric Surfaces

Let A be a 3 × 3 real symmetric matrix. The locus of the equation

   
x x
[x, y , z]A  y  + [b1 , b2 , b3 ]  y  + c = 0 (4)
z z

in three variables is called a quadric surface. We can carry out an

analysis similar to the one for quadratic forms in two variables to bring
(4) into standard form. Degenerate cases may arise. But the primary
cases are :

27/33
Equation Surface signs of eigenvalues of A

x2 y2 z2
+ + ellipsoid all three positive
a2 b2 c2
x2 y2 z
+ − =0 elliptic paraboloid two positive, one negative
a2 b2 c
x2 y2 z 2
+ − =0 elliptic cone two positive, one negative
a2 b2 c2
x2 y2 z 2
+ − =1 1-sheeted hyperboloid two positive, one negative
a2 b2 c2
x2 y2 z 2
− − =1 2-sheeted hyperboloid one positive, two negative
a2 b2 c2
x2 y2 z
− − =0 hyperbolic paraboloid one positive, one negative
a2 b2 c

28/33
Example
(1) 7x 2 + 7y 2 − 2z 2 + 20yz − 20zx − 2xy = 36.

29/33
Example
(1) 7x 2 + 7y 2 − 2z 2 + 20yz − 20zx − 2xy = 36.
The matrix form is:
   
7 −1 −10 x
[x, y , z]  −1 7 10   y  = 36.
−10 10 −2 z

29/33
Example
(1) 7x 2 + 7y 2 − 2z 2 + 20yz − 20zx − 2xy = 36.
The matrix form is:
   
7 −1 −10 x
[x, y , z]  −1 7 10   y  = 36.
−10 10 −2 z

Let A be the 3 × 3 matrix appearing in this equation. The eigenvalues

of A are λ1 = 6, λ2 = −12 and λ2 = 18.

29/33
Example
(1) 7x 2 + 7y 2 − 2z 2 + 20yz − 20zx − 2xy = 36.
The matrix form is:
   
7 −1 −10 x
[x, y , z]  −1 7 10   y  = 36.
−10 10 −2 z

Let A be the 3 × 3 matrix appearing in this equation. The eigenvalues

of A are λ1 = 6, λ2 = −12 and λ2 = 18. An orthonormal set of
eigenvectors is given by the column vectors of the orthogonal matrix
 1
√1 √1

√  
2 6 3 6 0 0
U =  √12 − √16 − √13  and U t AU =  0 −12 0  .
 
0 √2 − √1 0 0 18
6 3

29/33
Example
Consider the change of coordinates given by
   
x u
 y  = U  v .
z w

30/33
Example
Consider the change of coordinates given by
   
x u
 y  = U  v .
z w

This change of coordinates transforms the given equation into the form

6u 2 − 12v 2 + 18w 2 = 36

or
u2 v 2 w 2
− + =1
6 3 2
This is a hyperboloid of one sheet.

30/33
Example
(2) Consider the quadric

x 2 + y 2 + z 2 + 4yz − 4zx + 4xy = 27.

31/33
Example
(2) Consider the quadric

x 2 + y 2 + z 2 + 4yz − 4zx + 4xy = 27.

 
1 2 −2
The symmetric matrix for this is A =  2 1 2  . The
−2 2 1
eigenvalues of A are 3, 3, −3.

31/33
Example
(2) Consider the quadric

x 2 + y 2 + z 2 + 4yz − 4zx + 4xy = 27.

 
1 2 −2
The symmetric matrix for this is A =  2 1 2  . The
−2 2 1
eigenvalues of A are 3, 3, −3. Since A is diagonalizable the
eigenspace E(3) is 2-dimensional. We find a basis of E(3) first.

31/33
Example
(2) Consider the quadric

x 2 + y 2 + z 2 + 4yz − 4zx + 4xy = 27.

 
1 2 −2
The symmetric matrix for this is A =  2 1 2  . The
−2 2 1
eigenvalues of A are 3, 3, −3. Since A is diagonalizable the
eigenspace E(3) is 2-dimensional. We find a basis of E(3) first. The
eigenvectors for λ = 3 are solutions of
     
−2 2 −2 x 0
 2 −2 2   y  = 0 .

−2 2 −2 z 0

Hence we obtain the equation x − y + z = 0.

31/33
Example
(2) Consider the quadric

x 2 + y 2 + z 2 + 4yz − 4zx + 4xy = 27.

 
1 2 −2
The symmetric matrix for this is A =  2 1 2  . The
−2 2 1
eigenvalues of A are 3, 3, −3. Since A is diagonalizable the
eigenspace E(3) is 2-dimensional. We find a basis of E(3) first. The
eigenvectors for λ = 3 are solutions of
     
−2 2 −2 x 0
 2 −2 2   y  = 0 .

−2 2 −2 z 0

Hence we obtain the equation x − y + z = 0. Hence

E(3) = {(y − z, y , z) | y , z ∈ R}
= L({u = (0, 1, 1)t , u = (−1, 1, 2)t }). 31/33
Example
Now we apply Gram-Schmidt process to get an orthonormal basis of
E(3): r
1 1 t 2 1 1
v1 = (0, √ , √ ) and v2 = (− , − √ , √ ).
2 2 3 6 6

32/33
Example
Now we apply Gram-Schmidt process to get an orthonormal basis of
E(3): r
1 1 t 2 1 1
v1 = (0, √ , √ ) and v2 = (− , − √ , √ ).
2 2 3 6 6
√ √ √
A unit eigenvector for λ = −3 is v3 = (1/ 3, 1/ 3, 1/ 3). We know
that hv3 , v1 i = hv3 , v2 i = 0 since A is a symmetric matrix.

32/33
Example
Now we apply Gram-Schmidt process to get an orthonormal basis of
E(3): r
1 1 t 2 1 1
v1 = (0, √ , √ ) and v2 = (− , − √ , √ ).
2 2 3 6 6
√ √ √
A unit eigenvector for λ = −3 is v3 = (1/ 3, 1/ 3, 1/ 3). We know
that hv3 , v1 i = hv3 , v2 i = 0 since A is a symmetric matrix. The
orthogonal matrix for diagonalization is U = [v1 , v2 , v3 ] written column
wise. The quadric under the change of coordinates
   
x u
 y =U v 
z w

reduces to 3u 2 + 3v 2 − 3w 2 = 27.

32/33
Example
Now we apply Gram-Schmidt process to get an orthonormal basis of
E(3): r
1 1 t 2 1 1
v1 = (0, √ , √ ) and v2 = (− , − √ , √ ).
2 2 3 6 6
√ √ √
A unit eigenvector for λ = −3 is v3 = (1/ 3, 1/ 3, 1/ 3). We know
that hv3 , v1 i = hv3 , v2 i = 0 since A is a symmetric matrix. The
orthogonal matrix for diagonalization is U = [v1 , v2 , v3 ] written column
wise. The quadric under the change of coordinates
   
x u
 y =U v 
z w

reduces to 3u 2 + 3v 2 − 3w 2 = 27. This is a hyperboloid of one sheet.

32/33
Example

33/33