Title: On implicit-factorization block preconditioners
1On implicit-factorization block preconditioners
- Sue Dollar1,2, Nick Gould3,
- Wil Schilders2,4 and Andy Wathen1
1 Oxford University Computing Laboratory,
Oxford, UK 2 CASA, Technische Universiteit
Eindhoven, The Netherlands 3 Rutherford Appleton
Laboratory, Chilton, UK 4 Philips Research
Laboratory, Eindhoven, The Netherlands
2Summary of the talk
- Introduction
- Direct versus Iterative Methods
- Preconditioned Conjugate Gradient Method
- Constraint Preconditioners
- Implicit Constraint Preconditioners
- Numerical Examples
- Future work and conclusions
3Introduction
- Interested in solving structured linear systems
of equations -
- full rank
- positive semi-definite
4Example 1 Equality constrained minimization
- Interior-point sub-problem
-
-
- T (small) barrier weights
- First-order optimality
5Factory
6Example 2 Inequality constrained minimization
- Interior-point sub-problem
- C (small) barrier weights
- First-order optimality
7Direct vs. Iterative Methods
- Direct methods
- Gaussian elimination with a pivoting strategy
- Bunch-Parlett factorization
- Iterative methods
- Krylov subspace methods
- MINRES GMRES find solution of (1) within nm
iterations - CG may fail because of the indefiniteness of
system - Often advantageous to use a preconditioner P
8PCG
- Possible to use the preconditioned
conjugate-gradient method -
-
-
-
-
(2)
(Gould, Hribar, Nocedal)
9- Projected PCG
- Can perform iteration in the original (x, z)
space so long as preconditioner chosen carefully - Solve
- Iterate until convergence
-
- Solve
-
10Constraint Preconditioners
- The matrix P-1H has
- an eigenvalue at 1 with multiplicity 2m
- n-m eigenvalues which are defined by the
generalized eigenvalue problem
(Keller, Gould, Wathen)
11- The matrix P-1H has
- at least 2m-l eigenvalues at 1
- n-m eigenvalues which are defined by the
generalized eigenvalue problem - m eigenvalues defined by the generalized
eigenvalue problem - where wxy xz, subject to Y2BY1xy?0.
12CVXQP1_S m50, n100, Gdiag(A)
13CVXQP1_S m50, n100, Gdiag(A)
14Explicit vs. implicit constraint preconditioner
- Explicit constraint preconditioners choose G,
and then factorize - Implicit constraint preconditioners find
easily-invertible factors R and S so that - always holds
- Pick parts of R and S to match ( G) to parts
of A
15- Require that both R and S are easily block
invertible - - some sub-blocks should be zero
16Example 1 C0
- can recover any G
(Schilders)
17Example 2
Example 3
18Numerical Examples
CUTEr test set Case C0
Explicit GI Explicit GI Explicit GI Implicit G22I Implicit G22I Implicit G22I
name n m fact. iter. total fact. iter. total ratio
CVXQP1 10000 5000 1.73 2 2.98 0.066 27 1.47 0.49
CVXQP2 10000 2500 0.045 2 0.15 0.032 9 0.28 1.80
CVXQP3 10000 7500 3.11 2 6.09 0.085 16 1.35 0.22
POWELL20 10000 5000 0.12 1 0.19 0.040 314 6.52 33.0
PRIMALC1 239 9 0.0047 2 0.0092 0.0023 1 0.0046 0.50
PRIMALC2 238 7 0.0041 1 0.0071 0.0023 1 0.0042 0.58
QPNBOEI1 726 351 0.028 59 0.44 0.0044 13 0.060 0.13
QPNBOEI2 305 166 0.0071 55 0.11 0.0029 17 0.031 0.29
UBH1 9009 6000 0.11 9 0.50 0.049 1 0.16 0.33
26/40 problems solved faster (20 if take into
account perm time)
19Case CI
Explicit GI Explicit GI Explicit GI Implicit G22I Implicit G22I Implicit G22I
name n m fact. iter. total fact. iter. total ratio
CVXQP1 10000 5000 1.98 2 3.41 0.062 14 1.02 0.30
CVXQP2 10000 2500 0.048 2 0.16 0.030 11 0.33 2.10
CVXQP3 10000 7500 6.73 2 10.5 0.083 8 1.09 0.10
POWELL20 10000 5000 0.081 2 0.17 0.040 2 0.13 0.77
PRIMALC1 239 9 0.0045 1 0.0080 0.0024 1 0.0044 0.55
PRIMALC2 238 7 0.0037 1 0.0069 0.0021 1 0.0038 0.56
QPNBOEI1 726 351 0.029 26 0.27 0.0040 17 0.077 0.28
QPNBOEI2 305 166 0.0080 6 0.031 0.0027 14 0.032 1.03
UBH1 9009 6000 0.12 2 0.26 0.043 2 0.22 0.83
29/40 problems solved faster (25 if take into
account perm time)
20Conclusions
- New method for constructing preconditioners for
CG methods for a variety of important problems - Preconditioners
- implicitly respect crucial structure
- easy to apply
- flexible and capable of improving eigenvalue
clusters - Extend the class of problems for which CG is
applicable - Still under developmentbut will be available as
part of GALAHAD - Open questions
- How to pick basis B1 efficiently and so as to
improve eigenvalue clusters - How to approximate blocks of H in G
- What about Mr Greedy?
21(No Transcript)
22Conclusions
- New method for constructing preconditioners for
CG methods for a variety of important problems - Preconditioners
- implicitly respect crucial structure
- easy to apply
- flexible and capable of improving eigenvalue
clusters - Extend the class of problems for which CG is
applicable - Still under developmentbut will be available as
part of GALAHAD - Open questions
- How to pick basis B1 efficiently and so as to
improve eigenvalue clusters - How to approximate blocks of H in G