Math 416 - Abstract Linear Algebra

Fall 2011, section E1

Schur decomposition

Let us illustrate the algorithm to find a Schur decomposition, as in § 6.1, Theorem 1.1.

Example: Find a Schur decomposition of the matrix

 
7 −2
A= .
12 −3

Solution: First, we want an eigenvector of A. Let us find the eigenvalues:

 
7 − λ −2 
det(A − λI) =  
12 −3 − λ
= (7 − λ)(−3 − λ) + 24
= λ2 − 4λ − 21 + 24
= λ2 − 4λ + 3
= (λ − 1)(λ − 3).

The eigenvalues are λ = 1, 3. We could arbitrarily pick one of the two and find an eigenvector,
but while we’re at it, let’s find both:
   
6 −2 3 −1
λ = 1 : A − λI = A − I = ∼
12 −4 0 0
   
1 1 1
Take , normalized to 10
√ .
3 3
   
4 −2 2 −1
λ = 3 : A − λI = A − 3I = ∼
12 −6 0 0
   
1 1
Take , normalized to √15 .
2 2
 
1 1
In fact, let’s pick λ1 = 3 with normalized eigenvector u1 = √5 .
2
   
⊥ 1 ⊥ −2
We need to find an orthonormal basis {v2 } of Span{u1 } = Span{ } = Span{ }. Pick
  2 1
−2
v2 = √15 .
1
Now we express the transformation A in the new orthonormal basis {u1 , v2 }. Writing
 
  1 1 −2
U = u1 v2 = √
5 2 1

1
the matrix of A in the basis {u1 , v2 } is

[A]{u1 ,v2 } = U −1 [A]standard U

= U T [A]standard U
    
1 1 2 7 −2 1 1 −2
=√ √
5 −2 1 12 −3 5 2 1
  
1 31 −8 1 −2
=
5 −2 1 2 1
 
1 15 −70
=
5 0 5
 
3 −14
= .
0 1

Because A was 2 × 2, we can take u2 = v2 and the algorithm stops here. We have found the
Schur decomposition A = U T U ∗ where
 
3 −14
T =
0 1

is upper triangular and  

1 1 −2
U=√
5 2 1
is unitary.

Remark: The Schur decomposition is not unique, as there are choices involved in the algo-
rithm. However, the eigenvalues of A will always appear on the diagonal of T , since A is similar
to T .
The theorem does not guarantee that U and T will be real matrices, even if we start with a
real matrix A.

Example: Find a Schur decomposition of the matrix

 
1 1
A= .
−2 3

Solution: First, we want an eigenvector of A. Let us find the eigenvalues:

 
1 − λ 1 
det(A − λI) = 
−2 3 − λ
= (1 − λ)(3 − λ) + 2
= λ2 − 4λ + 5
p
4± 16 − 4(5) √
λ= = 2 ± −1 = 2 ± i.
2

2
Pick λ1 = 2 + i. (The choice doesn’t really matter since eigenvectors of a real matrix corre-
sponding to complex eigenvalues come in conjugate pairs.) Let us find an eigenvector:
   
−1 − i 1 1 + i −1
λ1 = 2 + i : A − λ1 I = A − (2 + i)I = ∼
−2 1−i 0 0
   
1 1 1
Take , normalized to u1 = √3 . Note that for complex matrices, the null space is
1+i 1+i
not orthogonal to the row space, but rather to the conjugate row space.
We need to find an orthonormal basis {v2 } of
 
⊥ 1
Span{u1 } = Span{ }⊥
1+i
 
= Null 1 1 − i
 
1−i
= Span{ }.
−1
 
1 1 − i
Take the normalized vector v2 = √3 .
−1
Now we express the transformation A in the new orthonormal basis {u1 , v2 }. Writing
 
  1 1 1−i
U = u1 v2 = √
3 1 + i −1
the matrix of A in the basis {u1 , v2 } is

[A]{u1 ,v2 } = U −1 [A]standard U

= U ∗ [A]standard U
    
1 1 1−i 1 1 1 1 1−i
=√ √
3 1 + i −1 −2 3 3 1 + i −1
  
1 −1 + 2i 4 − 3i 1 1−i
=
3 3 + i −2 + i 1 + i −1
 
1 6 + 3i −3 + 6i
=
3 0 6 − 3i
 
2 + i −1 + 2i
= .
0 2−i

Because A was 2 × 2, we can take u2 = v2 and the algorithm stops here. We have found the
Schur decomposition A = U T U ∗ where
 
2 + i −1 + 2i
T =
0 2−i
is upper triangular and  
1 1 1−i
U=√
3 1 + i −1
is unitary.

3
Conclusion: Computing a Schur decomposition by hand is annoying. In practice1 , knowing
the existence of a Schur decomposition is more useful than finding an explicit one.

1
for the purposes of pure mathematics