Unitary Diagonalization of Matrices

Here we take a different approach than Lang and diagonalize all matrices which can possibly be diago-
nalized using unitary matrices.
Theorem 1. The product of two unitary matrices is unitary.
Proof: Suppose Q and S are unitary, so Q−1 = Q∗ and S −1 = S ∗ . Then (QS)∗ = S ∗ Q∗ = S −1 Q−1 =
(QS)−1 so QS is unitary
Theorem 2. (Schur Lemma) If A is any square complex matrix then there is an upper triangular complex
matrix U and a unitary matrix S so that A = SU S ∗ = SU S −1 .
Proof: Let q1 be an eigenvector of A, which we may suppose has unit length. By the Gram-Schmidt process
we may chooseqi0 so that{q1 , q20 , . . . , qn0 } is an orthonormal basis. Let Q0 = [q1 q20 · · · qn0 ], then Q0 is unitary
λ1 ∗
and Q∗0 AQ0 = for some (n−1)×(n−1) matrix A1 . Likewise, we may find a unitary (n−1)×(n−1)
0 A1  
    λ1 ∗ ∗
λ2 ∗ 1 0
matrix Q1 so that Q∗1 A1 Q1 = . Then if S1 = Q0 we have S1∗ AS1 =  0 λ2 ∗ .
0 A2 0 Q1
 0 0 A2
Ik 0
Note that S1 is unitary by Theorem 2. Now continue in this fashion, letting Sk = Sk−1 , and we
0 Qk
∗ ∗
see that U = Sn ASn is upper triangular. Letting S = Sn we see that A = SU S .
Finally we characterize which matrices can be diagonalized by a unitary matrix. We say a matrix A is
normal if AA∗ = A∗ A.
Theorem 3. A matrix A is diagonalizable with a unitary matrix if and only if A is normal. In other words:
a) If A is normal there is a unitary matrix S so that S ∗ AS is diagonal.
b) If there is a unitary matrix S so that S ∗ AS is diagonal then A is normal.
Proof: Suppose A is normal. By Theorem 2 there is a unitary matrix S and an upper triangular U so that
A = SU S ∗ . Then

U U ∗ = S ∗ AS(S ∗ AS)∗ = S ∗ ASS ∗ A∗ S = S ∗ AA∗ S = S ∗ A∗ AS = S ∗ A∗ SS ∗ AS = U ∗ U

But if we let uij denote the ij-th entry of U then the upper left entry of U ∗ U is u11 u11 = |u11 |2 but the
upper left entry of U U ∗ is

u11 u11 + u12 u12 + · · · + u1n u1n = |u11 |2 + |u12 |2 + · · · + |u1n |2

Since this equals |u11 |2 and all summands are nonnegative real numbers we must have u12 = u13 = · · · =
u1n = 0. Similarly, looking at the second diagonal entry we see that u2j = 0 for all j > 2. Continuing in
this way we see that U must be diagonal. So we have shown that if A is normal, then it is diagonalizable
with a unitary matrix.
Now suppose that A is any matrix so that there is a unitary matrix S so that S ∗ AS = D is diagonal.
Note DD∗ = D∗ D. Then

AA∗ = SDS ∗ (SDS ∗ )∗ = SDS ∗ SD∗ S ∗ = SDD∗ S ∗ = SD∗ DS ∗ = SD∗ S ∗ SDS ∗ = A∗ A

Consequently, A is normal.
Examples of normal matrices are Hermitian matrices (A = A∗ ), skew Hermitian matrices (A = −A∗ )
and unitary matrices (A∗ = A−1 ) so all such matrices are diagonalizable.

1
 The Schur Lemma above needed to use a complex unitary matrix S. Note that A and U have the same
characteristic polynomial and hence the diagonal entries of U are the eigenvalues of A. So if A is a real
matrix and we want to find a real unitary matrix S so that tSAS is upper triangular, this is only possible
if all the eigenvalues of A are real. But a real matrix A often has nonreal eigenvalues. The following real
Schur Lemma shows what we can do in the real case.
But first a definition. For the purposes of these notes, we will say that a matrix U is 2 × 2 block upper
triangular if the entries uij of U satisfy:
a) uij = 0 if i > j + 1.
b) If ui+1,i 6= 0, then the adjacent entries below the diagonal are 0, i.e., ui+2,i+1 = 0 and ui,i−1 = 0.
In other words a matrix U is 2 × 2 block upper triangular if we can write U in block form as:
B
11 B12 ··· B1,k−1 B1k 
 0 B22 ··· B2,k−1 B2k 

U = .. .. .. .. .. 
 . . . . .


0 0 · · · Bk−1,k−1 Bk−1,k
 
0 0 ··· 0 Bkk

where each Bij is a 2 × 2 or 1 × 1 matrix.

Theorem 4. (real Schur Lemma) If A is any square real matrix then there is a 2 × 2 block upper triangular
real matrix U and a real unitary matrix S so that A = SU tS = SU S −1 . The eigenvalues of each 2 × 2 block
on the diagonal are a complex conjugate pair of non real eigenvalues of A.
Proof: The proof is similar to that of theorem 2. If A has a real eigenvalue then we let q1 be a real
eigenvector and proceed as in theorem 2. If A has no real eigenvalues, let a + bi be an eigenvalue and let
u + iv be a nonzero eigenvector with u and v real. Then

Au + iAv = A(u + iv) = (a + bi)(u + iv) = au − bv + i(bu + av)

Equating real and imaginary parts, we see that Au = au − bv and Av = bu + av. Note that u 6= 0 since
if u = 0 we would have 0 = Au = au − bv = −bv so v = 0 since b 6= 0. Likewise v 6= 0. We know that
v and u are linearly independent because if v = cu then Au = au − bcu = (a − bc)u so A would have a
real eigenvalue a − bc. We now let {q1 , q2 } be an orthonormal basis for the subspace generated by u and v,
for example q1 = u/||u|| and q2 = (v− < v, q1 > q1 )/||v− < v, q1 > q1 ||. Extend to  an orthonormal
 basis
B 1 C1
{q1 , q2 , . . . , qn } of Rn . Let Q0 = [q1 q2 · · · qn ], then Q0 is real unitary and Q∗0 AQ0 = for some
0 A1
2 × 2 matrix B1 , some (n − 2) × (n − 2) matrix A1 , and some 2 × (n − 2) matrix C1 . Now continue as in the
proof of theorem 2. It is not hard to see that the determinent of a block triangular matrix is the product of
the determinents of the diagonal blocks, and hence the characteristic polynomial of U is the product of the
characteristic polynomials of its diagonal blocks. Since U and A have the same characteristic polynomial,
each eigenvalue of a diagonal block of U is an eigenvalue of A. Eigenvalues of real matrices come in complex
conjugate pairs and the 2 × 2 blocks correspond to Note that if all the eigenvalues of A are real, then the
resulting U will be upper triangular.
If A is symmetric then, as shown in Lang, all its eigenvalues are real so we get tSAS = U is upper
triangular, so
t
U = tS tAS = tSAS = U
so U is diagonal.
If A is real unitary, then we saw that all eigenvalues λ have modulus 1, λλ = 1.
—More to come, but I’ll post this incomplete version now