Lecture 20

Spectral Theorem
Definition 1 (Orthogonal Matrix). A real square matrix is called orthogonal if AAT = I = AT A.

Definition 2 (Unitary Matrix). A complex square matrix is called unitary if AA∗ = I = A∗ A, where A∗
T
is the conjugate transpose of A, that is, A∗ = A .

Theorem 3. Let A be a unitary (real orthogonal) matrix. Then

(i) rows of A forms an orthonormal set;
(ii) columns of A forms an orthonormal set.

Remark 4. 1. P is orthogonal if and only if P T is orthogonal.

2. P is unitary if and only if P ∗ is unitary.
3. An orthogonal matrix (unitary) is invertible and its inverse is orthogonal (unitary).
4. Product of two orthogonal (unitary) matrices is orthogonal (unitary).

Theorem 5. The eigenvalues of a unitary matrix (an orthogonal matrix) has absolute value 1.

Proof: Let λ be an eigenvalue of a unitary matrix A. Then there exists a non-zero vector X such that
AX = λX. Thus, (AX)∗ = λ̄X ∗ ⇒ (AX)∗ (AX) = λ̄X ∗ (λX) ⇒ X ∗ A∗ AX = λλ̄X ∗ X. But A∗ A = I,
(1 − |λ|2 )X ∗ X = 0, i.e., |λ| = 1.

Definition 6. A complex square matrix A is called a Hermitian matrix if A = A∗ , where A∗ is the

T
conjugate transpose of A, that is, A∗ = A . A complex square matrix is called skew-Hermitian if A = −A∗ .

Theorem 7. 1. The eigenvalues of a Hermitian matrix (real symmetric matrix) are real.
2. The eigenvalues of a skew-Hermitian matrix (real skew-symmetric matrix) are either purely imaginary
or zero.

Proof: Let λ be an eigenvalue of a Hermitian matrix A. Then there exists a non-zero vector X ∈ Cn
such that AX = λX, multiplying both side by X ∗ , we get X ∗ AX = λX ∗ X. Taking conjugate transpose
both sides, we get (X ∗ AX)∗ = (λX ∗ X)∗ ⇒ X ∗ AX = λ̄X ∗ X. Thus we see that λX ∗ X = λ̄X ∗ X. Since
X 6= 0, X ∗ X = ||X||2 6= 0 so that λ = λ̄. For skew-Hermitian matrix, proceed in a similar way.

Theorem 8. Let A be a real symmetric matrix. Then eigenvectors of A corresponding to distinct

eigenvalues are orthogonal.

Proof: Let λ1 6= λ2 be two eigenvalues of A and v1 and v2 be corresponding eigenvectors respectively.

Then Av1 = λ1 v1 ⇒ v1T AT = λ1 v1T ⇒ v1T AT v2 = λ1 v1T v2 . Also (Av1 )T v2 = v1T AT v2 = v1T Av2 = λ2 v1T v2 .
Hence, (λ1 − λ2 )v1T v2 = 0, and λ1 6= λ2 so that v1T v2 = 0 = hv1 , v2 i ⇒ v1 ⊥ v2 .

Theorem 9. [Spectral Theorem for a real symmetric matrix] Let A be a real symmetric matrix.
Then there exists an orthogonal matrix P such that P T AP = D, where D is a diagonal matrix. In other
words, a real symmetric matrix is orthogonally diagonalizable.

Proof: The proof is by induction on order of the matrix. The result holds for n = 1. Suppose the
result holds for (n − 1) × (n − 1) symmetric matrix. Let A be a symmetric matrix of order n × n. Note

1
that A has real eigenvalues. Let λ ∈ R be one of the eigenvalue and 0 6= X ∈ Rn be a corresponding
eigenvector with norm 1, then AX = λX. Construct an orthonormal basis (by Gram-Schmidt process)
B = {v1 , v2 , v3 , . . . , vn }, where v1 = X and vi ∈ Rn . Construct a matrix P such that the i-th column of
P is vi . Then P is an orthogonal matrix.

Note that the matrix P −1 AP is symmetric and the first column of P −1 AP is given by P −1 AP (e
1 ), thus

−1 −1 −1 −1 λ 0
P A(P e1 ) = P AX = P λX = λe1 . Therefore, the matrix can be represented as P AP = ,
0 C
where C is a symmetric matrix of order (n − 1) × (n − 1). Hence, by induction hypothesis, C is similar
−1 T
to a diagonal
 matrix, say D, i.e., there is an orthogonal matrix Q such that Q CQ = Q CQ = D. Let
1 0
R=P . We claim that R is orthogonal and RT AR is diagonal.
0 Q
   
1 0
−1 −1 1 0
R = P = P T = RT , and
0 Q−1 0 QT
           
T 1 0 T 1 0 1 0 λ 0 1 0 λ 0 λ 0
R AR = P AP = = = .
0 QT 0 Q 0 QT 0 C 0 Q 0 QT CQ 0 D
Thus R is an orthogonal matrix such that RT AR is diagonal. Therefore, A is orthogonally diagonalizable.


Theorem 10. Converse of the above theorem is also true, i.e., if A ∈ Mn (R) is orthogonally diagonaliz-
able, then A is symmetric.

Proof: Let A be a matrix which is orthogonally diagonalizable. Then there is an orthogonal matrix P
s.t. P −1 AP = P T AP = D, equivalentely, A = P DP −1 = P DP T . This shows that AT = A. Hence
proved.

Example : Find an orthogonal matrix P and a diagonal matrix D such that P T AP = D, where
 
1 2 2
A = 2 1 2 .
2 2 1

The characteristic polynomial is (x + 1)2 (x − 5). The eigenvalues are 5, −1, −1. An eigenvector cor-
responding to λ = 5 is v1 = (1, 1, 1). The two independent eigenvectors corresponding to λ = −1 are
v2 = (−1, 0, 1) and v3 = (−1, 1, 0). Thus, B = {v1 , v2 , v3 } forms a basis of R3 . To find an orthonormal
basis, we apply Gram-Schmidt process on B. Thus
√
w1 = v1 , ||w1 || = 3, √
w2 = v2 (eigen vectors corresponding to distinct eigen values are orthogonal), ||w2 || = √ 2,
w3 = v3 − hv||w
3 ,w1 iw1
1 ||
2 − hv||w
3 ,w2 iw2
2 ||
2 = (−1, 1, 0) − 0(1, 1, 1) − 1(−1,0,1)
2
= (− 12 , 1, − 12 ), ||w3 || = 26
1
√1 √1

√
3
− 2
− 
5 0 0

√ 6
6 
Thus, P =  √13 0 and D = 0 −1 0  . Verify yourself that P T AP = D.

3 
√1 − √1 − √1 0 0 −1
3 3 6