You are on page 1of 40

Triangularization of matrix polynomials in their most

general form
Leo Taslaman
The University of Manchester
September 10, 2012
Joint work with Fran¸ coise Tisseur
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 1 / 18
Triangularization
(Schur form) For any square A, there exists an unitary Q such that
QAQ

is triangular.
(Generalized Schur form) For any square Aλ + B, there exist unitary
Q and Z such that Q(Aλ + B)Z

is triangular.
What can be said about triangularization for higher order matrix
polynomials A

λ

+ A
−1
λ
−1
+· · · + A
0
?
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 2 / 18
Triangularization
(Schur form) For any square A, there exists an unitary Q such that
QAQ

is triangular.
(Generalized Schur form) For any square Aλ + B, there exist unitary
Q and Z such that Q(Aλ + B)Z

is triangular.
What can be said about triangularization for higher order matrix
polynomials A

λ

+ A
−1
λ
−1
+· · · + A
0
?
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 2 / 18
Triangularization
(Schur form) For any square A, there exists an unitary Q such that
QAQ

is triangular.
(Generalized Schur form) For any square Aλ + B, there exist unitary
Q and Z such that Q(Aλ + B)Z

is triangular.
What can be said about triangularization for higher order matrix
polynomials A

λ

+ A
−1
λ
−1
+· · · + A
0
?
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 2 / 18
A link to the previous talk
I λ

+ A
−1
λ
−1
+· · · + A
0
Equivalence
−−−−−−−−→ I λ

+ T
−1
λ
−1
+· · · + T
0
. ¸¸ .
Triangular
¸
¸
_Linearization
¸
¸
¸
I
n

_
¸
¸
¸
¸
¸
_
−A
0
I −A
1
I −A
2
.
.
.
.
.
.
I −A
−1
_
¸
¸
¸
¸
¸
_
Similarity
−−−−−−−−→ I
n

_
¸
¸
¸
¸
¸
_
−T
0
I −T
1
I −T
2
.
.
.
.
.
.
I −T
−1
_
¸
¸
¸
¸
¸
_
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 3 / 18
Objectives for this talk
We will show that any square P(λ) over an algebraically closed field is
strongly equivalent to a triangular matrix polynomial of the same
degree.
Similarly, over R, we show strong equivalence to a quasi-triangular
matrix polynomial of the same degree.
Inverse problems.
A conjecture.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 4 / 18
Objectives for this talk
We will show that any square P(λ) over an algebraically closed field is
strongly equivalent to a triangular matrix polynomial of the same
degree.
Similarly, over R, we show strong equivalence to a quasi-triangular
matrix polynomial of the same degree.
Inverse problems.
A conjecture.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 4 / 18
Objectives for this talk
We will show that any square P(λ) over an algebraically closed field is
strongly equivalent to a triangular matrix polynomial of the same
degree.
Similarly, over R, we show strong equivalence to a quasi-triangular
matrix polynomial of the same degree.
Inverse problems.
A conjecture.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 4 / 18
Objectives for this talk
We will show that any square P(λ) over an algebraically closed field is
strongly equivalent to a triangular matrix polynomial of the same
degree.
Similarly, over R, we show strong equivalence to a quasi-triangular
matrix polynomial of the same degree.
Inverse problems.
A conjecture.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 4 / 18
Definitions
Consider an n ×n matrix polynomial over a field F, that is is a matrix
_
¸
_
p
11
(λ) . . . p
1n
(λ)
.
.
.
.
.
.
.
.
.
p
n1
(λ) . . . p
nn
(λ)
_
¸
_
, p
ij
(λ) ∈ F[λ],
or equivalently, a polynomial
P(λ) = A

λ

+ A
−1
λ
−1
+· · · + A
0
, A
i
∈ F
n×n
.
We assume that A

= 0 and is called the degree. The reversal is defined
as
revP(λ) = A
0
λ

+ A
1
λ
−1
+· · · + A

.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 5 / 18
Definitions
Consider an n ×n matrix polynomial over a field F, that is is a matrix
_
¸
_
p
11
(λ) . . . p
1n
(λ)
.
.
.
.
.
.
.
.
.
p
n1
(λ) . . . p
nn
(λ)
_
¸
_
, p
ij
(λ) ∈ F[λ],
or equivalently, a polynomial
P(λ) = A

λ

+ A
−1
λ
−1
+· · · + A
0
, A
i
∈ F
n×n
.
We assume that A

= 0 and is called the degree. The reversal is defined
as
revP(λ) = A
0
λ

+ A
1
λ
−1
+· · · + A

.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 5 / 18
Definitions
Consider an n ×n matrix polynomial over a field F, that is is a matrix
_
¸
_
p
11
(λ) . . . p
1n
(λ)
.
.
.
.
.
.
.
.
.
p
n1
(λ) . . . p
nn
(λ)
_
¸
_
, p
ij
(λ) ∈ F[λ],
or equivalently, a polynomial
P(λ) = A

λ

+ A
−1
λ
−1
+· · · + A
0
, A
i
∈ F
n×n
.
We assume that A

= 0 and is called the degree.
The reversal is defined
as
revP(λ) = A
0
λ

+ A
1
λ
−1
+· · · + A

.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 5 / 18
Definitions
Consider an n ×n matrix polynomial over a field F, that is is a matrix
_
¸
_
p
11
(λ) . . . p
1n
(λ)
.
.
.
.
.
.
.
.
.
p
n1
(λ) . . . p
nn
(λ)
_
¸
_
, p
ij
(λ) ∈ F[λ],
or equivalently, a polynomial
P(λ) = A

λ

+ A
−1
λ
−1
+· · · + A
0
, A
i
∈ F
n×n
.
We assume that A

= 0 and is called the degree. The reversal is defined
as
revP(λ) = A
0
λ

+ A
1
λ
−1
+· · · + A

.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 5 / 18
Smith form
Any P(λ) can be reduced to a unique Smith form by the elementary
transformations:
1
Permutation of rows/columns.
2
Multiply a row/column by a nonzero element in F.
3
Add one row/column to another row/column.
If P(λ) and Q(λ) have the same Smith form we say that they are
equivalent and write P(λ) ∼ Q(λ).
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 6 / 18
Smith form
Any P(λ) can be reduced to a unique Smith form by the elementary
transformations:
1
Permutation of rows/columns.
2
Multiply a row/column by a nonzero element in F.
3
Add one row/column to another row/column.
If P(λ) and Q(λ) have the same Smith form we say that they are
equivalent and write P(λ) ∼ Q(λ).
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 6 / 18
Smith form – a closer look
_
¸
¸
¸
_
d
1
.
.
.
0
r ,n−r
d
r
0
n−r ,r
0
n−r ,n−r
_
¸
¸
¸
_
, d
1
| · · · |d
r
are monic polynomials,
and d
i
= φ
α
1i
1
φ
α
2i
2
· · · φ
α
ki
k
. The φ
i
’s are distinct and irreducible over F[λ].
φ
α
ti
j
is a finite elementary divisor.
If F is algebraically closed, the φ
i
’s are linear and display the finite
eigenvalues. The α
ij
’s are the corresponding partial multiplicities.
If F = R the φ
i
’s are linear or quadratic.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 7 / 18
Smith form – a closer look
_
¸
¸
¸
_
d
1
.
.
.
0
r ,n−r
d
r
0
n−r ,r
0
n−r ,n−r
_
¸
¸
¸
_
, d
1
| · · · |d
r
are monic polynomials,
and d
i
= φ
α
1i
1
φ
α
2i
2
· · · φ
α
ki
k
. The φ
i
’s are distinct and irreducible over F[λ].
φ
α
ti
j
is a finite elementary divisor.
If F is algebraically closed, the φ
i
’s are linear and display the finite
eigenvalues. The α
ij
’s are the corresponding partial multiplicities.
If F = R the φ
i
’s are linear or quadratic.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 7 / 18
Smith form – a closer look
_
¸
¸
¸
_
d
1
.
.
.
0
r ,n−r
d
r
0
n−r ,r
0
n−r ,n−r
_
¸
¸
¸
_
, d
1
| · · · |d
r
are monic polynomials,
and d
i
= φ
α
1i
1
φ
α
2i
2
· · · φ
α
ki
k
. The φ
i
’s are distinct and irreducible over F[λ].
φ
α
ti
j
is a finite elementary divisor.
If F is algebraically closed, the φ
i
’s are linear and display the finite
eigenvalues. The α
ij
’s are the corresponding partial multiplicities.
If F = R the φ
i
’s are linear or quadratic.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 7 / 18
Smith form – a closer look
_
¸
¸
¸
_
d
1
.
.
.
0
r ,n−r
d
r
0
n−r ,r
0
n−r ,n−r
_
¸
¸
¸
_
, d
1
| · · · |d
r
are monic polynomials,
and d
i
= φ
α
1i
1
φ
α
2i
2
· · · φ
α
ki
k
. The φ
i
’s are distinct and irreducible over F[λ].
φ
α
ti
j
is a finite elementary divisor.
If F is algebraically closed, the φ
i
’s are linear and display the finite
eigenvalues. The α
ij
’s are the corresponding partial multiplicities.
If F = R the φ
i
’s are linear or quadratic.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 7 / 18
Elementary divisors at infinity and strong equivalence
The elementary divisors at infinity of P(λ) are the elementary divisors
at 0 of revP(λ).
If P(λ) ∼ Q(λ) and all the elementary divisors at infinity coincide, we
say that P(λ) and Q(λ) are strongly equivalent.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 8 / 18
Elementary divisors at infinity and strong equivalence
The elementary divisors at infinity of P(λ) are the elementary divisors
at 0 of revP(λ).
If P(λ) ∼ Q(λ) and all the elementary divisors at infinity coincide, we
say that P(λ) and Q(λ) are strongly equivalent.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 8 / 18
Example
Example
Consider
P(λ) =
_
_
1 0 0
0 λ −2 0
0 0 0
_
_
.
The reversal is
revP(λ) =
_
_
λ 0 0
0 1 −2λ 0
0 0 0
_
_

_
_
1 0 0
0 (λ −
1
2
)λ 0
0 0 0
_
_
so P(λ) has an elementary divisor at infinity.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 9 / 18
Another example
Example
Consider another matrix polynomial Q(λ) equivalent to P(λ)
Q(λ) =
_
_
1 0 λ
0 λ −2 0
0 0 0
_
_

_
_
1 0 0
0 λ −2 0
0 0 0
_
_
= P(λ).
The reversal is
revQ(λ) =
_
_
λ 0 1
0 1 −2λ 0
0 0 0
_
_

_
_
1 0 0
0 λ −
1
2
0
0 0 0
_
_
so Q(λ) has an no elementary divisors at infinity.
Conclusion: Elementary transformations preserve equivalence but not in
general strong equivalence.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 10 / 18
Another example
Example
Consider another matrix polynomial Q(λ) equivalent to P(λ)
Q(λ) =
_
_
1 0 λ
0 λ −2 0
0 0 0
_
_

_
_
1 0 0
0 λ −2 0
0 0 0
_
_
= P(λ).
The reversal is
revQ(λ) =
_
_
λ 0 1
0 1 −2λ 0
0 0 0
_
_

_
_
1 0 0
0 λ −
1
2
0
0 0 0
_
_
so Q(λ) has an no elementary divisors at infinity.
Conclusion: Elementary transformations preserve equivalence but not in
general strong equivalence.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 10 / 18
M¨ obius transformations of matrix polynomials
Theorem (Mackey, Mackey, Mehl and Mehrmann 2012)
Let P(λ) =

i =0
λ
i
A
i
∈ F[λ]
n×n
. If m(λ) =
aλ+b
cλ+d
is invertible, then
m(λ
0
) is an eigenvalue of P(λ) (∞ allowed) iff λ
0
is an eigenvalue of
Q(λ) = (cλ + d)

i =0
_
aλ + b
cλ + d
_
i
A
i
,
with the same partial multiplicity sequence.
Q(λ) has the same zero-structure as P(λ).
If we can triangularize any matrix polynomial with only finite
eigenvalues, then we can triangularize any matrix polynomial.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 11 / 18
M¨ obius transformations of matrix polynomials
Theorem (Mackey, Mackey, Mehl and Mehrmann 2012)
Let P(λ) =

i =0
λ
i
A
i
∈ F[λ]
n×n
. If m(λ) =
aλ+b
cλ+d
is invertible, then
m(λ
0
) is an eigenvalue of P(λ) (∞ allowed) iff λ
0
is an eigenvalue of
Q(λ) = (cλ + d)

i =0
_
aλ + b
cλ + d
_
i
A
i
,
with the same partial multiplicity sequence.
Q(λ) has the same zero-structure as P(λ).
If we can triangularize any matrix polynomial with only finite
eigenvalues, then we can triangularize any matrix polynomial.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 11 / 18
Unwinding the Smith form over an algebraically closed field
Consider P(λ) of degree with only finite eigenvalues.
We can “deflate” a degree polynomial from the Smith form:
_
¸
¸
¸
¸
¸
_
d
1
.
.
.
d
r −1
d
r
_
¸
¸
¸
¸
¸
_

_
¸
¸
¸
¸
¸
_
˜
d
1

.
.
.
.
.
.
˜
d
r −1

p
_
¸
¸
¸
¸
¸
_
where deg(p) = , and diag(
˜
d
1
, . . . ,
˜
d
n
) is in Smith form [Matrix
polynomials by GLR, Thm 1.7].
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 12 / 18
... and finally
T(λ) =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
t
1
∗ . . . ∗
.
.
.
.
.
.
.
.
.
.
.
.
t
k
∗ . . . ∗
× . . . ∗
.
.
.
.
.
.
×
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
where deg(t
i
) < for i = 1 : k, and all × on the diagonal are of degree .
T(λ) is triangular, of degree and equivalent to P(λ).
Problem: We may have added eigenvalues at infinite.
Have developed a technique to remove unwanted infinite eigenvalues.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 13 / 18
... and finally
T(λ) =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
t
1
∗ . . . ∗
.
.
.
.
.
.
.
.
.
.
.
.
t
k
∗ . . . ∗
× . . . ∗
.
.
.
.
.
.
×
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
where deg(t
i
) < for i = 1 : k, and all × on the diagonal are of degree .
T(λ) is triangular, of degree and equivalent to P(λ).
Problem: We may have added eigenvalues at infinite.
Have developed a technique to remove unwanted infinite eigenvalues.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 13 / 18
... and finally
T(λ) =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
t
1
∗ . . . ∗
.
.
.
.
.
.
.
.
.
.
.
.
t
k
∗ . . . ∗
× . . . ∗
.
.
.
.
.
.
×
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
where deg(t
i
) < for i = 1 : k, and all × on the diagonal are of degree .
T(λ) is triangular, of degree and equivalent to P(λ).
Problem: We may have added eigenvalues at infinite.
Have developed a technique to remove unwanted infinite eigenvalues.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 13 / 18
... and finally
T(λ) =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
t
1
∗ . . . ∗
.
.
.
.
.
.
.
.
.
.
.
.
t
k
∗ . . . ∗
× . . . ∗
.
.
.
.
.
.
×
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
where deg(t
i
) < for i = 1 : k, and all × on the diagonal are of degree .
T(λ) is triangular, of degree and equivalent to P(λ).
Problem: We may have added eigenvalues at infinite.
Have developed a technique to remove unwanted infinite eigenvalues.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 13 / 18
Quasi-triangularization over R
We use a real M¨obius transformation and deal with infinite
elementary divisors as in the algebraically closed case.
We cannot always deflate
_
¸
¸
¸
¸
¸
_
d
1
.
.
.
d
r −1
d
r
_
¸
¸
¸
¸
¸
_

_
¸
¸
¸
¸
¸
_
˜
d
1

.
.
.
.
.
.
˜
d
r −1

p
_
¸
¸
¸
¸
¸
_
where deg p = .
Theorem (T., Tisseur and Zaballa 2012)
If we cannot deflate a degree polynomial over R[λ], then we can
consecutive deflate two polynomias of degrees + 1 and −1 respectively.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 14 / 18
Quasi-triangularization over R
We use a real M¨obius transformation and deal with infinite
elementary divisors as in the algebraically closed case.
We cannot always deflate
_
¸
¸
¸
¸
¸
_
d
1
.
.
.
d
r −1
d
r
_
¸
¸
¸
¸
¸
_

_
¸
¸
¸
¸
¸
_
˜
d
1

.
.
.
.
.
.
˜
d
r −1

p
_
¸
¸
¸
¸
¸
_
where deg p = .
Theorem (T., Tisseur and Zaballa 2012)
If we cannot deflate a degree polynomial over R[λ], then we can
consecutive deflate two polynomias of degrees + 1 and −1 respectively.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 14 / 18
Quasi-triangularization over R
We use a real M¨obius transformation and deal with infinite
elementary divisors as in the algebraically closed case.
We cannot always deflate
_
¸
¸
¸
¸
¸
_
d
1
.
.
.
d
r −1
d
r
_
¸
¸
¸
¸
¸
_

_
¸
¸
¸
¸
¸
_
˜
d
1

.
.
.
.
.
.
˜
d
r −1

p
_
¸
¸
¸
¸
¸
_
where deg p = .
Theorem (T., Tisseur and Zaballa 2012)
If we cannot deflate a degree polynomial over R[λ], then we can
consecutive deflate two polynomias of degrees + 1 and −1 respectively.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 14 / 18
Forming 2 ×2-blocks on the diagonal
We get a triangular matrix polynomial with diagonal entries of degrees ,
−1 and + 1.
Consider a diagonal block where deg(d
i
) = −1 and deg(d
i +1
) = + 1
_
d
i

0 d
i +1
_

_
d
i
d
i
λ +∗
0 d
i +1
_

_
d
i
d
i
λ +∗
−d
i
λ q
_
. ¸¸ .
degree
where deg(q) ≤
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 15 / 18
Forming 2 ×2-blocks on the diagonal
We get a triangular matrix polynomial with diagonal entries of degrees ,
−1 and + 1.
Consider a diagonal block where deg(d
i
) = −1 and deg(d
i +1
) = + 1
_
d
i

0 d
i +1
_

_
d
i
d
i
λ +∗
0 d
i +1
_

_
d
i
d
i
λ +∗
−d
i
λ q
_
. ¸¸ .
degree
where deg(q) ≤
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 15 / 18
Inverse problems
We have build matrix polynomials from the Smith form having the
desired elementary divisors at infinity. This solves some inverse
polynomial eigenvalue problems.
Any Smith form such that nonreal elementary divisors comes in
complex conjugate pairs, is admissible by a real n ×n matrix
polynomial of minimal degree.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 16 / 18
Inverse problems
We have build matrix polynomials from the Smith form having the
desired elementary divisors at infinity. This solves some inverse
polynomial eigenvalue problems.
Any Smith form such that nonreal elementary divisors comes in
complex conjugate pairs, is admissible by a real n ×n matrix
polynomial of minimal degree.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 16 / 18
A conjecture
Nonreal elementary divisors of a regular self-adjoint matrix polynomial
A

λ

+ A
−1
λ
−1
+· · · + A
0
, A
i
= A

i
,
comes in complex conjugate pairs. Hence any regular self-adjoint matrix
polynomial is strongly equivalent to a real one.
Conjecture
Any regular real matrix polynomial is strongly equivalent to a regular
self-adjoint matrix polynomial, and vice versa.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 17 / 18
References
D. S. Mackey, N. Mackey, C. Mehl and V. Mehrmann, M¨obius
transformations of matrix polynomials. Technical report, 2012. In
preparation.
L. Taslaman, F. Tisseur and I. Zaballa Triangularizing matrix
polynomials. Technical report, 2012. In preparation.
Leo Taslaman The University of Manchester () Triangularization of matrix polynomials September 10, 2012 18 / 18