The Reduction to Hamiltonian Schur Form Explained David S. Watkins - - PowerPoint PPT Presentation

the reduction to hamiltonian schur form explained
SMART_READER_LITE
LIVE PREVIEW

The Reduction to Hamiltonian Schur Form Explained David S. Watkins - - PowerPoint PPT Presentation

The Reduction to Hamiltonian Schur Form Explained David S. Watkins watkins@math.wsu.edu Department of Mathematics Washington State University July 2006 p.1 Linear-Quadratic Gaussian Problem July 2006 p.2 Linear-Quadratic Gaussian


slide-1
SLIDE 1

The Reduction to Hamiltonian Schur Form Explained

David S. Watkins

watkins@math.wsu.edu

Department of Mathematics Washington State University

July 2006 – p.1

slide-2
SLIDE 2

Linear-Quadratic Gaussian Problem

July 2006 – p.2

slide-3
SLIDE 3

Linear-Quadratic Gaussian Problem

Optimal control theory

July 2006 – p.2

slide-4
SLIDE 4

Linear-Quadratic Gaussian Problem

Optimal control theory Solve algebraic Riccati equation

July 2006 – p.2

slide-5
SLIDE 5

Linear-Quadratic Gaussian Problem

Optimal control theory Solve algebraic Riccati equation Find stable invariant subspace of a Hamiltonian matrix

July 2006 – p.2

slide-6
SLIDE 6

Linear-Quadratic Gaussian Problem

Optimal control theory Solve algebraic Riccati equation Find stable invariant subspace of a Hamiltonian matrix Compute Hamiltonian Schur form (Paige/Van Loan 81)

July 2006 – p.2

slide-7
SLIDE 7

Linear-Quadratic Gaussian Problem

Optimal control theory Solve algebraic Riccati equation Find stable invariant subspace of a Hamiltonian matrix Compute Hamiltonian Schur form (Paige/Van Loan 81) Chu/Liu/Mehrmann 2004

July 2006 – p.2

slide-8
SLIDE 8

Hamiltonian Matrices

July 2006 – p.3

slide-9
SLIDE 9

Hamiltonian Matrices

H =

  • A

N K −AT

  • ∈ R2n×2n

NT = N KT = K

July 2006 – p.3

slide-10
SLIDE 10

Hamiltonian Matrices

H =

  • A

N K −AT

  • ∈ R2n×2n

NT = N KT = K

Hamiltonian spectrum

July 2006 – p.3

slide-11
SLIDE 11

Hamiltonian Matrices

H =

  • A

N K −AT

  • ∈ R2n×2n

NT = N KT = K

Hamiltonian spectrum assume no purely imaginary eigenvalues

July 2006 – p.3

slide-12
SLIDE 12

Hamiltonian Matrices

H =

  • A

N K −AT

  • ∈ R2n×2n

NT = N KT = K

Hamiltonian spectrum assume no purely imaginary eigenvalues stable invariant subspace

July 2006 – p.3

slide-13
SLIDE 13

Hamiltonian Matrices, continued

July 2006 – p.4

slide-14
SLIDE 14

Hamiltonian Matrices, continued

Hamiltonian Schur form:

  • T

N −T T

  • July 2006 – p.4
slide-15
SLIDE 15

Hamiltonian Matrices, continued

Hamiltonian Schur form:

  • T

N −T T

  • T is quasitriangular

July 2006 – p.4

slide-16
SLIDE 16

Hamiltonian Matrices, continued

Hamiltonian Schur form:

  • T

N −T T

  • T is quasitriangular

eigenvalues of T in left half plane

July 2006 – p.4

slide-17
SLIDE 17

Hamiltonian Matrices, continued

Hamiltonian Schur form:

  • T

N −T T

  • T is quasitriangular

eigenvalues of T in left half plane transforming matrix orthogonal and symplectic

July 2006 – p.4

slide-18
SLIDE 18

Hamiltonian Matrices, continued

Hamiltonian Schur form:

  • T

N −T T

  • T is quasitriangular

eigenvalues of T in left half plane transforming matrix orthogonal and symplectic This gives the stable invariant subspace.

July 2006 – p.4

slide-19
SLIDE 19

Hamiltonian Matrices, continued

Hamiltonian Schur form:

  • T

N −T T

  • T is quasitriangular

eigenvalues of T in left half plane transforming matrix orthogonal and symplectic This gives the stable invariant subspace. How can we compute it?

July 2006 – p.4

slide-20
SLIDE 20

Skew-Hamiltonian Matrices

July 2006 – p.5

slide-21
SLIDE 21

Skew-Hamiltonian Matrices

H Hamiltonian ⇒ H2 skew Hamiltonian

July 2006 – p.5

slide-22
SLIDE 22

Skew-Hamiltonian Matrices

H Hamiltonian ⇒ H2 skew Hamiltonian

skew-Hamiltonian Schur form:

  • B

M BT

  • July 2006 – p.5
slide-23
SLIDE 23

Skew-Hamiltonian Matrices

H Hamiltonian ⇒ H2 skew Hamiltonian

skew-Hamiltonian Schur form:

  • B

M BT

  • B is quasitriangular

July 2006 – p.5

slide-24
SLIDE 24

Skew-Hamiltonian Matrices

H Hamiltonian ⇒ H2 skew Hamiltonian

skew-Hamiltonian Schur form:

  • B

M BT

  • B is quasitriangular

easier than Hamiltonian

July 2006 – p.5

slide-25
SLIDE 25

Skew-Hamiltonian Matrices

H Hamiltonian ⇒ H2 skew Hamiltonian

skew-Hamiltonian Schur form:

  • B

M BT

  • B is quasitriangular

easier than Hamiltonian want to avoid squaring

July 2006 – p.5

slide-26
SLIDE 26

Skew-Hamiltonian Matrices

H Hamiltonian ⇒ H2 skew Hamiltonian

skew-Hamiltonian Schur form:

  • B

M BT

  • B is quasitriangular

easier than Hamiltonian want to avoid squaring symplectic URV decomposition

July 2006 – p.5

slide-27
SLIDE 27

Skew-Hamiltonian Matrices

H Hamiltonian ⇒ H2 skew Hamiltonian

skew-Hamiltonian Schur form:

  • B

M BT

  • B is quasitriangular

easier than Hamiltonian want to avoid squaring symplectic URV decomposition

H =

  • A

N K −AT

  • and

H2 =

  • B

M BT

  • July 2006 – p.5
slide-28
SLIDE 28

Assumptions

July 2006 – p.6

slide-29
SLIDE 29

Assumptions

all eigenvalues of H are real (time)

July 2006 – p.6

slide-30
SLIDE 30

Assumptions

all eigenvalues of H are real (time)

H2 =           ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.6

slide-31
SLIDE 31

Assumptions

all eigenvalues of H are real (time)

H2 =           ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

nongeneric cases ignored

July 2006 – p.6

slide-32
SLIDE 32

H2 =           ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.7

slide-33
SLIDE 33

H2 =           ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗           e1 is eigenvector of H2 but not of H

July 2006 – p.7

slide-34
SLIDE 34

H2 =           ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗           e1 is eigenvector of H2 but not of H

span{e1, He1} is invariant under H

July 2006 – p.7

slide-35
SLIDE 35

H2 =           ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗           e1 is eigenvector of H2 but not of H

span{e1, He1} is invariant under H

eigenvalues: ±λ1

July 2006 – p.7

slide-36
SLIDE 36

H2 =           ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗           e1 is eigenvector of H2 but not of H

span{e1, He1} is invariant under H

eigenvalues: ±λ1 Extract eigenvector x with eigenvalue λ1 < 0.

July 2006 – p.7

slide-37
SLIDE 37

Use eigenvector x.

July 2006 – p.8

slide-38
SLIDE 38

Use eigenvector x. Build orthogonal, symplectic transforming matrix with first column proportional to x.

July 2006 – p.8

slide-39
SLIDE 39

Use eigenvector x. Build orthogonal, symplectic transforming matrix with first column proportional to x.

Qe1 = αx QT x = α−1e1

July 2006 – p.8

slide-40
SLIDE 40

Use eigenvector x. Build orthogonal, symplectic transforming matrix with first column proportional to x.

Qe1 = αx QT x = α−1e1 ˆ H = QT HQ

July 2006 – p.8

slide-41
SLIDE 41

Use eigenvector x. Build orthogonal, symplectic transforming matrix with first column proportional to x.

Qe1 = αx QT x = α−1e1 ˆ H = QT HQ ˆ H =           λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.8

slide-42
SLIDE 42

Use eigenvector x. Build orthogonal, symplectic transforming matrix with first column proportional to x.

Qe1 = αx QT x = α−1e1 ˆ H = QT HQ ˆ H =           λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

Deflate?

July 2006 – p.8

slide-43
SLIDE 43

A Difficulty

July 2006 – p.9

slide-44
SLIDE 44

A Difficulty

What is the form of ˆ

H

2?

July 2006 – p.9

slide-45
SLIDE 45

A Difficulty

What is the form of ˆ

H

2?

Want ˆ

H

2 =

          ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.9

slide-46
SLIDE 46

A Difficulty

What is the form of ˆ

H

2?

Want ˆ

H

2 =

          ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗           Q must be built with care.

July 2006 – p.9

slide-47
SLIDE 47

Construction of QT

July 2006 – p.10

slide-48
SLIDE 48

Construction of QT

x =           ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.10

slide-49
SLIDE 49

Construction of QT

⇒ ⇒           ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.10

slide-50
SLIDE 50

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.10

slide-51
SLIDE 51

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗ ∗ ∗ ∗ ∗          

July 2006 – p.10

slide-52
SLIDE 52

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗ ∗ ∗ ∗ ∗          

July 2006 – p.10

slide-53
SLIDE 53

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗ ∗ ∗ ∗          

July 2006 – p.10

slide-54
SLIDE 54

Construction of QT

⇒ ⇒           ∗ ∗ ∗ ∗          

July 2006 – p.10

slide-55
SLIDE 55

Construction of QT

⇒ ⇒           ∗ ∗ ∗          

July 2006 – p.10

slide-56
SLIDE 56

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗ ∗ ∗          

July 2006 – p.10

slide-57
SLIDE 57

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗ ∗          

July 2006 – p.10

slide-58
SLIDE 58

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗ ∗          

July 2006 – p.10

slide-59
SLIDE 59

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗          

July 2006 – p.10

slide-60
SLIDE 60

Construction of QT

⇒ ⇒ ⇒ ⇒           ∗          

Done!

July 2006 – p.10

slide-61
SLIDE 61

Why it Works

July 2006 – p.11

slide-62
SLIDE 62

Why it Works

H2 stays in skew-Hamiltonian Schur form . . .

July 2006 – p.11

slide-63
SLIDE 63

Why it Works

H2 stays in skew-Hamiltonian Schur form . . .

. . . every step of the way.

July 2006 – p.11

slide-64
SLIDE 64

Why it Works

H2 stays in skew-Hamiltonian Schur form . . .

. . . every step of the way.

Hx = λ1x

July 2006 – p.11

slide-65
SLIDE 65

Why it Works

H2 stays in skew-Hamiltonian Schur form . . .

. . . every step of the way.

Hx = λ1x H2x = λ2

1x

July 2006 – p.11

slide-66
SLIDE 66

Why it Works

H2 stays in skew-Hamiltonian Schur form . . .

. . . every step of the way.

Hx = λ1x H2x = λ2

1x

relationships preserved throughout the transformation

July 2006 – p.11

slide-67
SLIDE 67

Close up

July 2006 – p.12

slide-68
SLIDE 68

Close up

H2x = xλ2

1

July 2006 – p.12

slide-69
SLIDE 69

Close up

H2x = xλ2

1

          λ2

1

∗ ∗ ∗ ∗ λ2

2

∗ ∗ ∗ λ2

3

∗ ∗ λ2

1

∗ λ2

2

∗ ∗ λ2

3

                    ∗ ∗ ∗ ∗ ∗ ∗           =           ∗ ∗ ∗ ∗ ∗ ∗           λ2

1

July 2006 – p.12

slide-70
SLIDE 70

Close up

H2x = xλ2

1

          λ2

1

∗ ∗ ∗ ∗ λ2

2

∗ ∗ ∗ λ2

3

∗ ∗ λ2

1

∗ λ2

2

∗ ∗ λ2

3

                    ∗ ∗ ∗ ∗ ∗ ∗           =           ∗ ∗ ∗ ∗ ∗ ∗           λ2

1

July 2006 – p.12

slide-71
SLIDE 71

   λ2

1

∗ λ2

2

∗ ∗ λ2

3

      ∗ ∗ ∗    =    ∗ ∗ ∗    λ2

1

July 2006 – p.13

slide-72
SLIDE 72

⇒ ⇒    λ2

1

∗ λ2

2

∗ ∗ λ2

3

      ∗ ∗ ∗    =    ∗ ∗ ∗    λ2

1

July 2006 – p.13

slide-73
SLIDE 73

⇒ ⇒    ∗ ∗ ∗ ∗ ∗ ∗ λ2

3

      ∗ ∗    =    ∗ ∗    λ2

1

July 2006 – p.13

slide-74
SLIDE 74

⇒ ⇒    ∗ ∗ ∗ ∗ ∗ λ2

3

      ∗ ∗    =    ∗ ∗    λ2

1

July 2006 – p.13

slide-75
SLIDE 75

⇒ ⇒    ∗ ∗ λ2

1

∗ ∗ λ2

3

      ∗ ∗    =    ∗ ∗    λ2

1

July 2006 – p.13

slide-76
SLIDE 76

⇒ ⇒    λ2

2

∗ λ2

1

∗ ∗ λ2

3

      ∗ ∗    =    ∗ ∗    λ2

1

July 2006 – p.13

slide-77
SLIDE 77

⇒ ⇒    λ2

2

∗ λ2

1

∗ ∗ λ2

3

      ∗ ∗    =    ∗ ∗    λ2

1

July 2006 – p.13

slide-78
SLIDE 78

⇒ ⇒    λ2

2

∗ ∗ ∗ ∗ ∗ ∗       ∗    =    ∗    λ2

1

July 2006 – p.13

slide-79
SLIDE 79

⇒ ⇒    λ2

2

∗ ∗ ∗ ∗ ∗       ∗    =    ∗    λ2

1

July 2006 – p.13

slide-80
SLIDE 80

⇒ ⇒    λ2

2

∗ ∗ ∗ ∗ λ2

1

      ∗    =    ∗    λ2

1

July 2006 – p.13

slide-81
SLIDE 81

⇒ ⇒    λ2

2

∗ λ2

3

∗ ∗ λ2

1

      ∗    =    ∗    λ2

1

July 2006 – p.13

slide-82
SLIDE 82

   λ2

2

∗ λ2

3

∗ ∗ λ2

1

      ∗    =    ∗    λ2

1

July 2006 – p.13

slide-83
SLIDE 83

          λ2

2

∗ ∗ ∗ ∗ λ2

3

∗ ∗ ∗ λ2

1

∗ ∗ λ2

2

∗ λ2

3

∗ ∗ λ2

1

                    ∗ ∗ ∗ ∗           =           ∗ ∗ ∗ ∗           λ2

1

July 2006 – p.14

slide-84
SLIDE 84

⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ λ2

3

∗ ∗ ∗ λ2

1

∗ ∗ λ2

2

∗ λ2

3

∗ ∗ λ2

1

                    ∗ ∗ ∗ ∗           =           ∗ ∗ ∗ ∗           λ2

1

July 2006 – p.14

slide-85
SLIDE 85

⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ λ2

3

∗ ∗ ∗ λ2

1

∗ ∗ λ2

2

∗ λ2

3

∗ ∗ λ2

1

                    ∗ ∗ ∗ ∗           =           ∗ ∗ ∗ ∗           λ2

1

July 2006 – p.14

slide-86
SLIDE 86

⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ λ2

3

∗ ∗ ∗ λ2

1

∗ ∗ λ2

2

∗ λ2

3

∗ ∗ λ2

1

                    ∗ ∗ ∗           =           ∗ ∗ ∗           λ2

1

July 2006 – p.14

slide-87
SLIDE 87

⇒ ⇒ ⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ λ2

3

∗ ∗ ∗ λ2

1

∗ ∗ λ2

2

∗ λ2

3

∗ ∗ λ2

1

                    ∗ ∗ ∗           =           ∗ ∗ ∗           λ2

1

July 2006 – p.14

slide-88
SLIDE 88

⇒ ⇒ ⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ λ2

3

∗ ∗ ∗ ∗ λ2

1

∗ ∗ λ2

2

∗ λ2

3

∗ ∗ ∗ λ2

1

                    ∗ ∗ ∗           =           ∗ ∗ ∗           λ2

1

July 2006 – p.14

slide-89
SLIDE 89

⇒ ⇒ ⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ λ2

1

∗ ∗ ∗ λ2

3

∗ ∗ λ2

2

∗ λ2

1

∗ ∗ λ2

3

                    ∗ ∗           =           ∗ ∗           λ2

1

July 2006 – p.14

slide-90
SLIDE 90

⇒ ⇒ ⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ λ2

1

∗ ∗ ∗ λ2

3

∗ ∗ λ2

2

∗ λ2

1

∗ ∗ λ2

3

                    ∗ ∗           =           ∗ ∗           λ2

1

July 2006 – p.14

slide-91
SLIDE 91

⇒ ⇒ ⇒ ⇒           λ2

2

∗ ∗ ∗ ∗ ∗ λ2

1

∗ ∗ ∗ λ2

3

∗ ∗ λ2

2

∗ ∗ λ2

1

∗ ∗ λ2

3

                    ∗ ∗           =           ∗ ∗           λ2

1

July 2006 – p.14

slide-92
SLIDE 92

⇒ ⇒ ⇒ ⇒           λ2

1

∗ ∗ ∗ ∗ λ2

2

∗ ∗ ∗ λ2

3

∗ ∗ λ2

1

∗ λ2

2

∗ ∗ λ2

3

                    ∗           =           ∗           λ2

1

July 2006 – p.14

slide-93
SLIDE 93

          λ2

1

∗ ∗ ∗ ∗ λ2

2

∗ ∗ ∗ λ2

3

∗ ∗ λ2

1

∗ λ2

2

∗ ∗ λ2

3

                    ∗           =           ∗           λ2

1

July 2006 – p.14

slide-94
SLIDE 94

          λ2

1

∗ ∗ ∗ ∗ λ2

2

∗ ∗ ∗ λ2

3

∗ ∗ λ2

1

∗ λ2

2

∗ ∗ λ2

3

                    ∗           =           ∗           λ2

1

It looks like we’re back where we started,

July 2006 – p.15

slide-95
SLIDE 95

          λ2

1

∗ ∗ ∗ ∗ λ2

2

∗ ∗ ∗ λ2

3

∗ ∗ λ2

1

∗ λ2

2

∗ ∗ λ2

3

                    ∗           =           ∗           λ2

1

It looks like we’re back where we started, but . . .

July 2006 – p.15

slide-96
SLIDE 96

H =           λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

July 2006 – p.16

slide-97
SLIDE 97

H =           λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −λ1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗          

Now we can deflate.

July 2006 – p.16

slide-98
SLIDE 98

In Conclusion

July 2006 – p.17

slide-99
SLIDE 99

In Conclusion

By keeping H2 in skew-Hamiltonian Schur form

July 2006 – p.17

slide-100
SLIDE 100

In Conclusion

By keeping H2 in skew-Hamiltonian Schur form every step of the way

July 2006 – p.17

slide-101
SLIDE 101

In Conclusion

By keeping H2 in skew-Hamiltonian Schur form every step of the way we can transform H to Hamiltonian Schur form.

July 2006 – p.17

slide-102
SLIDE 102

In Conclusion

By keeping H2 in skew-Hamiltonian Schur form every step of the way we can transform H to Hamiltonian Schur form. This work will appear in ETNA.

July 2006 – p.17

slide-103
SLIDE 103

In Conclusion

By keeping H2 in skew-Hamiltonian Schur form every step of the way we can transform H to Hamiltonian Schur form. This work will appear in ETNA. Download from my website. (Google David S Watkins)

July 2006 – p.17

slide-104
SLIDE 104

In Conclusion

By keeping H2 in skew-Hamiltonian Schur form every step of the way we can transform H to Hamiltonian Schur form. This work will appear in ETNA. Download from my website. (Google David S Watkins)

Thank you for your attention.

July 2006 – p.17