core-chasing algorithms for the eigenvalue problem · leonardo robol raf vandebril david s. watkins...

Core-Chasing Algorithmsfor the Eigenvalue Problem

David S. Watkins

Department of MathematicsWashington State University

HHXX, Virginia Tech, June 20, 2017

David S. Watkins Core-Chasing Algorithms

Our International Research Group

Collaborators:

Jared Aurentz

Thomas Mach

Leonardo Robol

Raf Vandebril

Today’s Topic

The matrix eigenvalue problem

A ∈ Cn×n

Find the eigenvalues (. . . vectors, invariant subspaces)

Possible structures

unitaryunitary-plus-rank-one (companion matrix)unitary-plus-low-rank. . . or no special structure

Today’s Topic

A ∈ Cn×n

Possible structures

Today’s Topic

A ∈ Cn×n

Possible structures

Today’s Topic

A ∈ Cn×n

Possible structures

Today’s Topic

A ∈ Cn×n

Possible structures

Today’s Topic

A ∈ Cn×n

Possible structures

unitaryunitary-plus-rank-one (companion matrix)unitary-plus-low-rank

. . . or no special structure

Today’s Topic

A ∈ Cn×n

Possible structures

John Francis

photo: Frank Uhlig, 2009

invented the winning algorithm in 1959.

implicitly shifted QR algorithm

Our algorithms are all variants of this.

John Francis

Francis’s algorithm . . .

. . . is a bulge chasing algorithm.

We turn it into a core chasing algorithm.

Instead of chasing bulges, we chase core transformations.

Core Transformations

What is a core transformation?

It’s a unitary matrix, and

it’s essentially 2× 2

1× ×× ×

Ex: Givens rotator, reflector, . . .

We just wanted a generic term.

1× ×× ×

× × ×× ×× ×

× × ×× ×0 ×

Abbreviated notation

�� × =

1× ×× ×

× × ×× ×× ×

× × ×× ×0 ×

�� × =

1× ×× ×

× × ×× ×× ×

× × ×× ×0 ×

�� × =

Hessenberg QR decomposition

×××××

�� ×××××

=××××

�� ×

××××

= ×××

��

��×××××

��

×××××

��

×××××

��

×××××

Now invert the core transformations.

×××××

��

Our algorithms operate on the matrix in QR decomposed form.

A = QR =

��

This is not inefficient.

We apply Francis’s algorithm to this factored form.

Operating on Core Transformations

Fusion

� �� ⇒ ��

Turnover

� ��

�� ⇔

[ ∗ ∗ ∗∗ ∗ ∗∗ ∗ ∗

]⇔ �

��

Turnover as a shift-through operation

��

Turnover as a shift-through operation : Abbreviated notation

��

Turnover as a shift-through operation : Abbreviated notation

��

Passing a core transformation through a triangular matrix

∗ ∗ ∗ ∗∗ ∗ ∗∗ ∗∗

�� ⇔

∗ ∗ ∗ ∗∗ ∗ ∗+ ∗ ∗

⇔ ��

∗ ∗ ∗ ∗∗ ∗ ∗∗ ∗∗

Cost is O(n) flops.

Passing a core transformation through a triangular matrix

Abbreviated Notation:

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Core Chasing

��

Flop Count

O(n3) total flops

O(n2) storage

about the same as for standard Francis iteration.

Flop Count

O(n3) total flops

O(n2) storage

about the same as for standard Francis iteration.

Advantages

Are there any advantages?

superior deflation procedure

some structured cases

Advantages

Deflation

Standard deflation criterion:

×××××

Set aj+1,j to zero if

|aj+1,j | < u (|aj ,j |+ |aj+1,j+1 |).

(u is unit roundoff.)

Deflation

Standard deflation criterion:

×××××

Set aj+1,j to zero if

|aj+1,j | < u (|aj ,j |+ |aj+1,j+1 |).

Deflation

Our deflation criterion:

��

cj −sjsj c j

Set sj to zero if |sj | < u.

Deflation

Our deflation criterion:

��

cj −sjsj c j

Set sj to zero if |sj | < u.

Deflation

Both criteria are normwise backward stable.

How does this affect eigenvalues?

Change in λ depends on condition number κ(λ).

Standard result: λ is perturbed to µ, where

|λ− µ | ≤ u κ(λ) ‖A‖+ O(u2).

This holds for both deflation criteria.

Deflation

|λ− µ | ≤ u κ(λ) ‖A‖+ O(u2).

Deflation

|λ− µ | ≤ u κ(λ) ‖A‖+ O(u2).

Deflation

|λ− µ | ≤ u κ(λ) ‖A‖+ O(u2).

Deflation

|λ− µ | ≤ u κ(λ) ‖A‖+ O(u2).

Deflation

But our criterion does better:

Theorem (Mach and Vandebril (2014) )

|λ− µ | ≤ u κ(λ) |λ |+ O(u2).

Relative perturbation in each λ is tiny.

This does not hold for standard deflation criterion.

Deflation

|λ− µ | ≤ u κ(λ) |λ |+ O(u2).

Deflation

|λ− µ | ≤ u κ(λ) |λ |+ O(u2).

Deflation

|λ− µ | ≤ u κ(λ) |λ |+ O(u2).

Deflation

Fun Example:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

These eigenvalues are well conditioned.

Standard criterion deflates to[1 20 ε

Eigenvalues are µ1 = 1 and µ2 = ε.

Small eigenvalue is off by 200%.

Deflation

Fun Example:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Deflation

Fun Example:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Deflation

Fun Example:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Deflation

Fun Example:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Deflation

Fun Example:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Deflation

Fun Example:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Deflation

Example, continued:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Our criterion:

A = QR ≈[

1 −εε 1

] [1 20 −ε

Deflates to [1 00 1

] [1 20 −ε

[1 20 −ε

Eigenvalues are µ1 = 1 and µ2 = −ε.Both eigenvalues are accurate.

Deflation

Example, continued:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Our criterion:

A = QR ≈[

1 −εε 1

] [1 20 −ε

Deflates to [1 00 1

] [1 20 −ε

[1 20 −ε

Deflation

Example, continued:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Our criterion:

A = QR ≈[

1 −εε 1

] [1 20 −ε

Deflates to [1 00 1

] [1 20 −ε

[1 20 −ε

Deflation

Example, continued:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Our criterion:

A = QR ≈[

1 −εε 1

] [1 20 −ε

Deflates to [1 00 1

] [1 20 −ε

[1 20 −ε

Deflation

Example, continued:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Our criterion:

A = QR ≈[

1 −εε 1

] [1 20 −ε

Deflates to [1 00 1

] [1 20 −ε

[1 20 −ε

Eigenvalues are µ1 = 1 and µ2 = −ε.

Both eigenvalues are accurate.

Deflation

Example, continued:

[1 2ε ε

](0 < ε < u)

λ1 = 1 + 2ε+ O(ε2) λ2 = −ε+ O(ε2)

Our criterion:

A = QR ≈[

1 −εε 1

] [1 20 −ε

Deflates to [1 00 1

] [1 20 −ε

[1 20 −ε

Exploitation of Structure

Structures we can exploit

unitary

companion matrix (unitary-plus-rank-one)

unitary-plus-low-rank

unitary

Unitary Case

A = QR =

��

Unitary Case

A = QR =

��

Unitary Case

A = QR =

��

Unitary Case

A = QR =

��

Cost is O(n) flops per iteration, O(n2) flops total.

Storage requirement is O(n).

Gragg (1986)

Ammar, Reichel, M. Stewart, Bunse-Gerstner, Elsner, He, W,. . .

Unitary Case

A = QR =

��

Cost is O(n) flops per iteration,

O(n2) flops total.

Gragg (1986)

Unitary Case

A = QR =

��

Gragg (1986)

Unitary Case

A = QR =

��

Gragg (1986)

Unitary Case

A = QR =

��

Gragg (1986)

Unitary Case

A = QR =

��

Gragg (1986)

Companion Case

p(x) = xn + an−1xn−1 + an−2x

n−2 + · · ·+ a0 = 0

monic polynomial

companion matrix

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

. . . get the zeros of p by computing the eigenvalues.

MATLAB’s roots command

Companion matrix is unitary-plus-rank-one.

Companion Case

p(x) = xn + an−1xn−1 + an−2x

n−2 + · · ·+ a0 = 0

monic polynomial

companion matrix

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

Companion Case

p(x) = xn + an−1xn−1 + an−2x

n−2 + · · ·+ a0 = 0

monic polynomial

companion matrix

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

Companion Case

p(x) = xn + an−1xn−1 + an−2x

n−2 + · · ·+ a0 = 0

monic polynomial

companion matrix

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

Companion Case

p(x) = xn + an−1xn−1 + an−2x

n−2 + · · ·+ a0 = 0

monic polynomial

companion matrix

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

Cost of solving companion eigenvalue problem

If structure not exploited:

O(n2) storage, O(n3) flopsFrancis’s algorithm

If structure exploited:

O(n) storage, O(n2) flopsseveral methods proposeddata-sparse representation + Francis’s algorithmOurs is fastest . . .. . . and we can prove backward stability.I spoke about this at the previous Householder symposium.

O(n) storage, O(n2) flops

several methods proposeddata-sparse representation + Francis’s algorithmOurs is fastest . . .. . . and we can prove backward stability.I spoke about this at the previous Householder symposium.

O(n) storage, O(n2) flopsseveral methods proposeddata-sparse representation + Francis’s algorithm

Ours is fastest . . .. . . and we can prove backward stability.I spoke about this at the previous Householder symposium.

O(n) storage, O(n2) flopsseveral methods proposeddata-sparse representation + Francis’s algorithmOurs is fastest . . .

. . . and we can prove backward stability.I spoke about this at the previous Householder symposium.

O(n) storage, O(n2) flopsseveral methods proposeddata-sparse representation + Francis’s algorithmOurs is fastest . . .. . . and we can prove backward stability.

I spoke about this at the previous Householder symposium.

Representation of R

We store the QR decomposed form.

A = QR =

��

1 0 · · · −a1

1 −a2. . .

...−a0

This is unitary-plus-rank-one.

How do we store it?

Representation of R

A = QR =

��

1 0 · · · −a1

1 −a2. . .

...−a0

How do we store it?

Representation of R

A = QR =

��

1 0 · · · −a1

1 −a2. . .

...−a0

How do we store it?

Representation of R

A = QR =

��

1 0 · · · −a1

1 −a2. . .

...−a0

How do we store it?

Representation of R

Add a row and column to R.

1 0 · · · −a1 0

1 −a2 0. . .

......

−a0 1

0 0 · · · 0 0

This is still unitary-plus-rank-one.

Representation of R

1 0 · · · −a1 0

1 −a2 0. . .

......

−a0 1

0 0 · · · 0 0

Representation of R

1 0 · · · −a1 0

1 −a2 0. . .

......

−a0 1

0 0 · · · 0 0

Representation of R

1 0 · · · 0 0

1 0 0. . .

......

0 0 · · · 1 0

0 0 · · · −a1 0

0 −a2 0. . .

......

−a0 0

0 0 · · · −1 0

Representation of R

C ∗n · · ·C ∗

1 (B1 · · ·Bn + e1yT )

��

+ · · ·

. . . and we don’t have to store the rank-one part!

This helps with backward stability.

Storage is O(n).

Representation of R

R = C ∗n · · ·C ∗

1 (B1 · · ·Bn + e1yT )

��

+ · · ·

Storage is O(n).

Representation of R

R = C ∗n · · ·C ∗

1 (B1 · · ·Bn + e1yT )

��

+ · · ·

Storage is O(n).

Representation of R

R = C ∗n · · ·C ∗

1 (B1 · · ·Bn + e1yT )

��

+ · · ·

Storage is O(n).

Representation of R

R = C ∗n · · ·C ∗

1 (B1 · · ·Bn + e1yT )

��

+ · · ·

Storage is O(n).

Representation of R

R = C ∗n · · ·C ∗

1 (B1 · · ·Bn + e1yT )

��

+ · · ·

Storage is O(n).

Passing a core transformation through R

��

C ∗n · · · C ∗

��

B1 · · · Bn

��

Cost: O(1) flops instead of O(n).

��

C ∗n · · · C ∗

��

B1 · · · Bn

��

C ∗n · · · C ∗

��

B1 · · · Bn

��

Other things we can do

We can also handle

generalized eigenvalue problem

companion pencil

matrix polynomial eigenvalue problems

L. Robol talk at 2 pm

generalizations of Hessenberg form

and more.

Monograph in progress (130+ pp.)

We can also handle

companion pencil

and more.

We can also handle

companion pencil

and more.

We can also handle

companion pencil

and more.

We can also handle

companion pencil

and more.

We can also handle

companion pencil

and more.

The Companion Pencil

p(x) = a0 + a1x + · · ·+ anxn

(not monic)

Divide by an, or . . .

companion pencil:

1. . .

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

We can handle this too (for a price),

This should be superior in some situations.

e.g. if an is tiny.

p(x) = a0 + a1x + · · ·+ anxn (not monic)

companion pencil:

1. . .

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

e.g. if an is tiny.

Divide by an,

or . . .

companion pencil:

1. . .

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

e.g. if an is tiny.

companion pencil:

1. . .

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

e.g. if an is tiny.

companion pencil:

1. . .

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

e.g. if an is tiny.

companion pencil:

1. . .

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

e.g. if an is tiny.

companion pencil:

1. . .

0 · · · 0 −a01 0 · · · 0 −a1

1. . .

......

. . . 0 −an−2

1 −an−1

e.g. if an is tiny.

Backward Stability

All of these algorithms are normwise backward stable.

This is “obvious” because we work only with unitarytransformations,

but it took us a while to write down a correct proof.

For details see . . .

Jared L. Aurentz, Thomas Mach, Raf Vandebril, and David S.Watkins, Fast and backward stable computation of roots ofpolynomials, SIAM J. Matrix Anal. Appl., 36 (2015), pp.942–973.

Co-winner of SIAM Best Paper Prize (2017).

Jared L. Aurentz, Thomas Mach, Leonardo Robol, RafVandebril, and David S. Watkins, Roots of polynomials: ontwisted QR methods for companion matrices and pencils,arXiv:1611.02435, currently undergoing a complete rewrite.

Backward Stability

Jared L. Aurentz, Thomas Mach, Leonardo Robol, RafVandebril, and David S. Watkins, Roots of polynomials: ontwisted QR methods for companion matrices and pencils,arXiv:1611.02435,

currently undergoing a complete rewrite.

Backward Stability

Backward Stability Odyssey

Presentation at Householder 2014

First written attempt (horrible)

Second attempt was much better (2015 paper) . . .

. . . but there was one one more thing!

Corrected in companion pencil paper. We also exploited thestructure of the backward error to get a better result.Rejected!

Search for examples.

Take a closer look.

backward error on pencil vs. polynomial coefficients

monic vs. scaled polynomial

Our analysis keeps getting better.

Stay tuned for the revised paper.

Take a closer look.

Corrected in companion pencil paper.

We also exploited thestructure of the backward error to get a better result.Rejected!

Take a closer look.

Corrected in companion pencil paper. We also exploited thestructure of the backward error to get a better result.

Rejected!

Take a closer look.

Nice Picture

‖a‖2

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

‖a−

our code

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

102‖a−

LAPACK balanced

Our code is not just faster, it is also more accurate!

Thank you for your attention.

Nice Picture

‖a‖2

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

‖a−

our code

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

102‖a−

LAPACK balanced

Our code is not just faster,

it is also more accurate!

Nice Picture

‖a‖2

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

‖a−

our code

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

102‖a−

LAPACK balanced

Nice Picture

‖a‖2

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

‖a−

our code

‖a‖22

101 103 105 10710−18

10−14

10−10

10−6

10−2

102‖a−

LAPACK balanced

core-chasing algorithms for the eigenvalue problem · leonardo robol raf vandebril david s. watkins...

Documents