solution of indefinite linear systems · s. lungten page 36 references [1] w. h. a. schilders,...

38
S. Lungten PAGE 0 PAGE 0 S. Lungten Promotor: W. H. A. Schilders Supervisor: J. M. L. Maubach SOLUTION OF INDEFINITE LINEAR SYSTEMS Center for Analysis, Scientific Computing and Applications (CASA) S. Lungten

Upload: others

Post on 12-Jul-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 0 PAGE 0

S. Lungten

Promotor: W. H. A. Schilders

Supervisor: J. M. L. Maubach

SOLUTION OF INDEFINITE LINEAR

SYSTEMS

Center for Analysis, Scientific Computing and

Applications (CASA)

S. Lungten

Page 2: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

PAGE 1

Outline

Introduction

Factorization of indefinite matrices

Numerical experiments

Conclusion and future work

S. Lungten

• Previous methods

• Current method

Page 3: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 2

Introduction

known as saddle point problems:

• is an symmetric and positive (semi)

definite.

To solve symmetric indefinite linear systems of the form

• is an symmetric and positive definite matrix,

• is an matrix of rank with ,

Page 4: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 3

• mixed finite element methods

• regularized, weighted least squares

• discretization of PDEs

• constraint optimization

• Stokes

• electric circuits and networks

• economic models

Introduction

Applications leading to saddle point problems:

The saddle point matrix appear in two forms:

Case 1.

Case 2.

Page 5: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 4

Introduction

• Direct solvers

• Iterative solvers

• Preconditioning techniques

Solution methods:

• existence

• sparsity permutation

• stability

Each method has issues with one of the following:

Page 6: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 5

Introduction

For example (in terms of sparsity), consider a matrix

,

and its permuted form:

.

Page 7: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 6

Introduction

The decomposition: gives:

while gives:

.

Page 8: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 7

PART I

Saddle point matrices

with the case

Page 9: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 8

where is an upper triangular and is an

matrix.

Solution method

I. Transform into as follows:

.

• is a permutation or an orthogonal matrix

• is congruent to

• is equivalent to such that

Page 10: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 9

Solution method

III. Factorize .

II. Split into a block 3 by 3 structure [S., 2009]:

.

IV. Solve by using the factors of where

Page 11: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 10

Existence of factorization (previous result)

Schilders [S., 2009] showed the factorization of the form :

exists, i.e.,

Page 12: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 11

because of the existence of :

This existence proof requires to be an

upper triangular form.

Existence of factorization (previous result)

.

Page 13: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 12

We do the following transformation:

such that where is an

lower triangular matrix.

Current approach : Transformation

Aim: sparse block factors

Page 14: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 13

Method relies on:

• Crout decomposition:

Current approach

• Micro-block factorization [Maubach and S., 2012]

(applied for the case upper triangular form of ).

-has computational efficiency equivalent to that of

Cholesky for a symmetric positive definite matrix .

Page 15: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 14

Micro-block partitioning

Algorithm for the factorization

Micro-block factorization

Back permutation

Page 16: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 15

Block partitioning: Example ( )

is macro-block partitioned:

Define a permutation matrix as in [S., 2009]:

Page 17: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 16

which leads to the micro-block partitioned matrix,

with micro-blocks of order 1 and 2.

Block partitioning: Example ( )

Page 18: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 17

leads to:

Micro-block factorization: eg ( )

Page 19: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 18

Back permuting and give:

Induced macro-block factors: eg ( )

Page 20: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 19

Macro-block factorization: Existence

For general and ,

Page 21: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 20

Micro-block partitioning

Algorithm for the factorization

Micro-block factorization

Back permutation

Page 22: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

PAGE 21 S. Lungten

Define,

Comparison with Schilders’

Eliminating and from and gives another

form, which resembles Schilders’ form.

Then,

where .

Page 23: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

PAGE 22 S. Lungten

Comparison with Schilders’

Differences between the non-trivial blocks:

Similarity between the non-trivial blocks:

Both and are determined from

, where is the basis of null space of .

Page 24: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 23

current through

resistors

injected current

through nodes

Resistor network modeling [Rommes and Schilders, 2010]:

Numerical experiments

voltage

across nodes

diagonal matrix

(resistance values

of resistors)

incidence matrix

• full rank

• (entries:

at most 2 nonzero in

each column )

Page 25: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

PAGE 24 S. Lungten

Two visual representations of graphs related to resistor

networks:

Numerical experiments

Page 26: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

PAGE 25 S. Lungten

Saddle point matrix representation of the graphs of

and :

Numerical experiments

Page 27: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

PAGE 26 S. Lungten

Transformation: of

Schilders’ , nz = 22371 Current , nz = 22371

Numerical experiments

Page 28: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 27

Macro-block factors of of :

Schilders’ , Current ,

Numerical experiments

Page 29: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 28

Schilders’ , nz = 23526 Current , nz = 23526

Numerical experiments

Transformation: of

Page 30: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 29

Schilders’ , Current ,

Numerical experiments

Macro-block factors of of :

Page 31: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 30

PART II

Saddle point matrices

with the case

Page 32: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 31

Regularized saddle point systems

Consider a regularized saddle point matrix:

S. Lungten PAGE 31

Consider for the case is a diagonal matrix.

• is an symmetric and positive (semi)

definite.

• is an symmetric and positive definite matrix,

• is an matrix of rank with ,

Page 33: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 32

Transformation of to

Due to :

• is congruent to

• is congruent to

• is equivalent to such that

, where is an

upper triangular matrix.

Page 34: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 33

Macro-block factorization of

The macro-block factorization is of the form:

This factorization exists and has its blocks with these

shapes, which is shown in [J.M.L. Maubach and S., “Micro- and

macro-block factorizations for regularized saddle point problems”,

submitted]

Page 35: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

S. Lungten PAGE 34

In particular,

where

Macro-block factorization of

Page 36: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

• The presented factorizations are all exact, but

can be used as a basis for preconditioning.

• Especially the method with lower triangular is

attractive for preconditiong (topic of further

research).

• More numerical experiments from various

applications.

PAGE 35 S. Lungten

Concluding remarks and future work

Page 37: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

PAGE 36 S. Lungten

References

[1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ

decomposition for the linear constraints, Linear Algebra and Applications,

pp. 381-395, 431 (2009).

[2] J. M. L. Maubach, W. H. A. Schilders, Micro- and macro-block factorizations for

regularized saddle point systems, Technical report, Center for Analysis, Scientific

computing and Applications, Eindhoven University of Technology, April 2012.

[3] J. Rommes and W. H. A. Schilders, Efficient Methods for Large Resistor

Networks, IEEE Transactions on Computer-Aided Design of Integrated Circuits

and Systems, 29 (2010), 28-39.

[4] S. Lungten, J. M. L. Maubach, W. H. A. Schilders, Sparse inverse incidence

matrices for Schilders’ factorization applied to resistor network modeling,

submitted to NACO, Dec. 2013.

Page 38: SOLUTION OF INDEFINITE LINEAR SYSTEMS · S. Lungten PAGE 36 References [1] W. H. A. Schilders, Solution of indefinite linear systems using an LQ decomposition for the linear constraints,

THANK YOU

PAGE 37 S. Lungten