Barret templates
1 like•990 views
This document provides an electronic version of the second edition of the Templates book, which describes templates (building blocks) for iterative methods used to solve systems of linear equations. The book is divided into chapters covering stationary and nonstationary iterative methods, preconditioning techniques, related computational issues like parallelism, and remaining topics. It includes summaries of various iterative methods like Jacobi, Gauss-Seidel, conjugate gradient, and GMRES as well as preconditioning approaches like incomplete factorizations. The document acknowledges funding sources and provides lists of symbols, figures, and contents for the book.
![2 CHAPTER 1. INTRODUCTION
the traditional goals of mathematical software, we believe that our approach will go a long way in
providing such a valuable service.
1.1 Why Use Templates?
Templates offer three significant advantages. First, templates are general and reusable. Thus, they
can simplify ports to diverse machines. This feature is important given the diversity of parallel
architectures.
Second, templates exploit the expertise of two distinct groups. The expert numerical analyst
creates a template reflecting in-depth knowledge of a specific numerical technique. The compu-
tational scientist then provides “value-added” capability to the general template description, cus-
tomizing it for specific contexts or applications needs.
And third, templates are not language specific. Rather, they are displayed in an Algol-like
structure, which is readily translatable into the target language such as FORTRAN (with the use
of the Basic Linear Algebra Subprograms, or BLAS, whenever possible) and C. By using these
familiar styles, we believe that the users will have trust in the algorithms. We also hope that users
will gain a better understanding of numerical techniques and parallel programming.
For each template, we provide some or all of the following:
• a mathematical description of the flow of the iteration;
• discussion of convergence and stopping criteria;
• suggestions for applying a method to special matrix types (e.g., banded systems);
• advice for tuning (for example, which preconditioners are applicable and which are not);
• tips on parallel implementations; and
• hints as to when to use a method, and why.
For each of the templates, the following can be obtained via electronic mail.
• a MATLAB implementation based on dense matrices;
• a FORTRAN-77 program with calls to BLAS1
.
See Appendix A for details.
1.2 What Methods Are Covered?
Many iterative methods have been developed and it is impossible to cover them all. We chose
the methods below either because they illustrate the historical development of iterative methods,
or because they represent the current state-of-the-art for solving large sparse linear systems. The
methods we discuss are:
1. Jacobi
2. Gauss-Seidel
3. Successive Over-Relaxation (SOR)
4. Symmetric Successive Over-Relaxation (SSOR)
1For a discussion of BLAS as building blocks, see [68, 69, 70, 143] and LAPACK routines [3]. Also, see Appendix B.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-12-320.jpg)
![10 CHAPTER 2. ITERATIVE METHODS
Choose an initial guess x(0)
to the solution x.
for k = 1, 2, . . .
for i = 1, 2, . . . , n
σ = 0
for j = 1, 2, . . . , i − 1
σ = σ + ai,jx
(k)
j
end
for j = i + 1, . . . , n
σ = σ + ai,jx
(k−1)
j
end
σ = (bi − σ)/ai,i
x
(k)
i = x
(k−1)
i + ω(σ − x
(k−1)
i )
end
check convergence; continue if necessary
end
Figure 2.3: The SOR Method
and the computed Gauss-Seidel iterate successively for each component:
x
(k)
i = ω¯x
(k)
i + (1 − ω)x
(k−1)
i
(where ¯x denotes a Gauss-Seidel iterate, and ω is the extrapolation factor). The idea is to choose a
value for ω that will accelerate the rate of convergence of the iterates to the solution.
In matrix terms, the SOR algorithm can be written as follows:
x(k)
= (D − ωL)−1
(ωU + (1 − ω)D)x(k−1)
+ ω(D − ωL)−1
b. (2.7)
The pseudocode for the SOR algorithm is given in Figure 2.3.
Choosing the Value of ω
If ω = 1, the SOR method simplifies to the Gauss-Seidel method. A theorem due to Kahan [129]
shows that SOR fails to converge if ω is outside the interval (0, 2). Though technically the term
underrelaxation should be used when 0 < ω < 1, for convenience the term overrelaxation is now
used for any value of ω ∈ (0, 2).
In general, it is not possible to compute in advance the value of ω that is optimal with respect
to the rate of convergence of SOR. Even when it is possible to compute the optimal value for ω, the
expense of such computation is usually prohibitive. Frequently, some heuristic estimate is used,
such as ω = 2 − O(h) where h is the mesh spacing of the discretization of the underlying physical
domain.
If the coefficient matrix A is symmetric and positive definite, the SOR iteration is guaranteed to
converge for any value of ω between 0 and 2, though the choice of ω can significantly affect the rate
at which the SOR iteration converges. Sophisticated implementations of the SOR algorithm (such
as that found in ITPACK [139]) employ adaptive parameter estimation schemes to try to home in
on the appropriate value of ω by estimating the rate at which the iteration is converging.
For coefficient matrices of a special class called consistently ordered with property A (see
Young [215]), which includes certain orderings of matrices arising from the discretization of el-
liptic PDEs, there is a direct relationship between the spectra of the Jacobi and SOR iteration](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-20-320.jpg)
![2.2. STATIONARY ITERATIVE METHODS 11
matrices. In principle, given the spectral radius ρ of the Jacobi iteration matrix, one can determine
a priori the theoretically optimal value of ω for SOR:
ωopt =
2
1 + 1 − ρ2
. (2.8)
This is seldom done, since calculating the spectral radius of the Jacobi matrix requires an imprac-
tical amount of computation. However, relatively inexpensive rough estimates of ρ (for example,
from the power method, see Golub and Van Loan [108, p. 351]) can yield reasonable estimates for
the optimal value of ω.
2.2.4 The Symmetric Successive Overrelaxation Method
If we assume that the coefficient matrix A is symmetric, then the Symmetric Successive Overre-
laxation method, or SSOR, combines two SOR sweeps together in such a way that the resulting
iteration matrix is similar to a symmetric matrix. Specifically, the first SOR sweep is carried out as
in (2.7), but in the second sweep the unknowns are updated in the reverse order. That is, SSOR is
a forward SOR sweep followed by a backward SOR sweep. The similarity of the SSOR iteration
matrix to a symmetric matrix permits the application of SSOR as a preconditioner for other iter-
ative schemes for symmetric matrices. Indeed, this is the primary motivation for SSOR since its
convergence rate, with an optimal value of ω, is usually slower than the convergence rate of SOR
with optimal ω (see Young [215, page 462]). For details on using SSOR as a preconditioner, see
Chapter 3.
In matrix terms, the SSOR iteration can be expressed as follows:
x(k)
= B1B2x(k−1)
+ ω(2 − ω)(D − ωU)−1
D(D − ωL)−1
b, (2.9)
where
B1 = (D − ωU)−1
(ωL + (1 − ω)D),
and
B2 = (D − ωL)−1
(ωU + (1 − ω)D).
Note that B2 is simply the iteration matrix for SOR from (2.7), and that B1 is the same, but with
the roles of L and U reversed.
The pseudocode for the SSOR algorithm is given in Figure 2.4.
2.2.5 Notes and References
The modern treatment of iterative methods dates back to the relaxation method of Southwell [192].
This was the precursor to the SOR method, though the order in which approximations to the un-
knowns were relaxed varied during the computation. Specifically, the next unknown was chosen
based upon estimates of the location of the largest error in the current approximation. Because
of this, Southwell’s relaxation method was considered impractical for automated computing. It is
interesting to note that the introduction of multiple-instruction, multiple data-stream (MIMD) paral-
lel computers has rekindled an interest in so-called asynchronous, or chaotic iterative methods (see
Chazan and Miranker [53], Baudet [30], and Elkin [91]), which are closely related to Southwell’s
original relaxation method. In chaotic methods, the order of relaxation is unconstrained, thereby
eliminating costly synchronization of the processors, though the effect on convergence is difficult
to predict.
The notion of accelerating the convergence of an iterative method by extrapolation predates
the development of SOR. Indeed, Southwell used overrelaxation to accelerate the convergence
of his original relaxation method. More recently, the ad hoc SOR method, in which a different](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-21-320.jpg)
![12 CHAPTER 2. ITERATIVE METHODS
Choose an initial guess x(0)
to the solution x.
Let x( 1
2 )
= x(0)
.
for k = 1, 2, . . .
for i = 1, 2, . . . , n
σ = 0
for j = 1, 2, . . . , i − 1
σ = σ + ai,jx
(k− 1
2 )
j
end
for j = i + 1, . . . , n
σ = σ + ai,jx
(k−1)
j
end
σ = (bi − σ)/ai,i
x
(k− 1
2 )
i = x
(k−1)
i + ω(σ − x
(k−1)
i )
end
for i = n, n − 1, . . . , 1
σ = 0
for j = 1, 2, . . . , i − 1
σ = σ + ai,jx
(k− 1
2 )
j
end
for j = i + 1, . . . , n
σ = σ + ai,jx
(k)
j
end
x
(k)
i = x
(k− 1
2 )
i + ω(σ − x
(k− 1
2 )
i )
end
check convergence; continue if necessary
end
Figure 2.4: The SSOR Method
relaxation factor ω is used for updating each variable, has given impressive results for some classes
of problems (see Ehrlich [82]).
The three main references for the theory of stationary iterative methods are Varga [209],
Young [215] and Hageman and Young [119]. The reader is referred to these books (and the
references therein) for further details concerning the methods described in this section.
2.3 Nonstationary Iterative Methods
Nonstationary methods differ from stationary methods in that the computations involve informa-
tion that changes at each iteration. Typically, constants are computed by taking inner products of
residuals or other vectors arising from the iterative method.
2.3.1 Conjugate Gradient Method (CG)
The Conjugate Gradient method is an effective method for symmetric positive definite systems. It
is the oldest and best known of the nonstationary methods discussed here. The method proceeds by
generating vector sequences of iterates (i.e., successsive approximations to the solution), residuals](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-22-320.jpg)
![14 CHAPTER 2. ITERATIVE METHODS
solution of Ax = b. This minimum is guaranteed to exist in general only if A is symmetric positive
definite. The preconditioned version of the method uses a different subspace for constructing the
iterates, but it satisfies the same minimization property, although over this different subspace. It
requires in addition that the preconditioner M is symmetric and positive definite.
The above minimization of the error is equivalent to the residuals r(i)
= b − Ax(i)
being M−1
orthogonal (that is, r(i)T
M−1
r(j)
= 0 if i = j). Since for symmetric A an orthogonal basis for the
Krylov subspace span{r(0)
, . . . , Ai−1
r(0)
} can be constructed with only three-term recurrences,
such a recurrence also suffices for generating the residuals. In the Conjugate Gradient method two
coupled two-term recurrences are used; one that updates residuals using a search direction vector,
and one updating the search direction with a newly computed residual. This makes the Conjugate
Gradient Method quite attractive computationally.
There is a close relationship between the Conjugate Gradient method and the Lanczos method
for determining eigensystems, since both are based on the construction of an orthogonal basis for
the Krylov subspace, and a similarity transformation of the coefficient matrix to tridiagonal form.
The coefficients computed during the CG iteration then arise from the LU factorization of this
tridiagonal matrix. From the CG iteration one can reconstruct the Lanczos process, and vice versa;
see Paige and Saunders [167] and Golub and Van Loan [108, §10.2.6]. This relationship can be
exploited to obtain relevant information about the eigensystem of the (preconditioned) matrix A;
see §5.1.
Convergence
Accurate predictions of the convergence of iterative methods are difficult to make, but useful
bounds can often be obtained. For the Conjugate Gradient method, the error can be bounded in
terms of the spectral condition number κ2 of the matrix M−1
A. (Recall that if λmax and λmin are
the largest and smallest eigenvalues of a symmetric positive definite matrix B, then the spectral
condition number of B is κ2(B) = λmax(B)/λmin(B)). If ˆx is the exact solution of the linear
system Ax = b, with symmetric positive definite matrix A, then for CG with symmetric positive
definite preconditioner M, it can be shown that
x(i)
− ˆx A ≤ 2αi
x(0)
− ˆx A (2.12)
where α = (
√
κ2 − 1)/(
√
κ2 + 1) (see Golub and Van Loan [108, §10.2.8], and Kaniel [130]),
and y 2
A ≡ (y, Ay). From this relation we see that the number of iterations to reach a relative
reduction of in the error is proportional to
√
κ2.
In some cases, practical application of the above error bound is straightforward. For example,
elliptic second order partial differential equations typically give rise to coefficient matrices A with
κ2(A) = O(h−2
) (where h is the discretization mesh width), independent of the order of the
finite elements or differences used, and of the number of space dimensions of the problem (see for
instance Axelsson and Barker [14, §5.5]). Thus, without preconditioning, we expect a number of
iterations proportional to h−1
for the Conjugate Gradient method.
Other results concerning the behavior of the Conjugate Gradient algorithm have been obtained.
If the extremal eigenvalues of the matrix M−1
A are well separated, then one often observes so-
called superlinear convergence (see Concus, Golub and O’Leary [57]); that is, convergence at a
rate that increases per iteration. This phenomenon is explained by the fact that CG tends to elimi-
nate components of the error in the direction of eigenvectors associated with extremal eigenvalues
first. After these have been eliminated, the method proceeds as if these eigenvalues did not exist
in the given system, i.e., the convergence rate depends on a reduced system with a (much) smaller
condition number (for an analysis of this, see Van der Sluis and Van der Vorst [197]). The effective-
ness of the preconditioner in reducing the condition number and in separating extremal eigenvalues
can be deduced by studying the approximated eigenvalues of the related Lanczos process.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-24-320.jpg)
![2.3. NONSTATIONARY ITERATIVE METHODS 15
Implementation
The Conjugate Gradient method involves one matrix-vector product, three vector updates, and two
inner products per iteration. Some slight computational variants exist that have the same structure
(see Reid [178]). Variants that cluster the inner products, a favorable property on parallel machines,
are discussed in §4.4.
For a discussion of the Conjugate Gradient method on vector and shared memory computers,
see Dongarra, et al. [70, 165]. For discussions of the method for more general parallel architectures
see Demmel, Heath and Van der Vorst [66] and Ortega [165], and the references therein.
Further references
A more formal presentation of CG, as well as many theoretical properties, can be found in the
textbook by Hackbusch [117]. Shorter presentations are given in Axelsson and Barker [14] and
Golub and Van Loan [108]. An overview of papers published in the first 25 years of existence of
the method is given in Golub and O’Leary [107].
2.3.2 MINRES and SYMMLQ
The Conjugate Gradient method can be viewed as a special variant of the Lanczos method (see §5.1)
for positive definite symmetric systems. The MINRES and SYMMLQ methods are variants that
can be applied to symmetric indefinite systems.
The vector sequences in the Conjugate Gradient method correspond to a factorization of a tridi-
agonal matrix similar to the coefficient matrix. Therefore, a breakdown of the algorithm can occur
corresponding to a zero pivot if the matrix is indefinite. Furthermore, for indefinite matrices the
minimization property of the Conjugate Gradient method is no longer well-defined. The MINRES
and SYMMLQ methods are variants of the CG method that avoid the LU-factorization and do not
suffer from breakdown. MINRES minimizes the residual in the 2-norm. SYMMLQ solves the
projected system, but does not minimize anything (it keeps the residual orthogonal to all previous
ones). The convergence behavior of Conjugate Gradients and MINRES for indefinite systems was
analyzed by Paige, Parlett, and Van der Vorst [166].
Theory
When A is not positive definite, but symmetric, we can still construct an orthogonal basis for the
Krylov subspace by three term recurrence relations. Eliminating the search directions in equa-
tions (2.10) and (2.11) gives a recurrence
Ar(i)
= r(i+1)
ti+1,i + r(i)
ti,i + r(i−1)
ti−1,i,
which can be written in matrix form as
ARi = Ri+1
¯Ti,
where ¯Ti is an (i + 1) × i tridiagonal matrix
¯Ti =
← i →
...
...
...
...
...
...
...
...
...
...
...
.
↑
i + 1
↓](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-25-320.jpg)
![16 CHAPTER 2. ITERATIVE METHODS
In this case we have the problem that (·, ·)A no longer defines an inner product. However we can
still try to minimize the residual in the 2-norm by obtaining
x(i)
∈ {r(0)
, Ar(0)
, . . . , Ai−1
r(0)
}, x(i)
= Ri ¯y
that minimizes
Ax(i)
− b 2 = ARi ¯y − b 2 = Ri+1
¯Tiy − b 2.
Now we exploit the fact that if Di+1 ≡ diag( r(0)
2, r(1)
2, ..., r(i)
2), then Ri+1D−1
i+1 is an
orthonormal transformation with respect to the current Krylov subspace:
Ax(i)
− b 2 = Di+1
¯Tiy − r(0)
2e(1)
2
and this final expression can simply be seen as a minimum norm least squares problem.
The element in the (i + 1, i) position of ¯Ti can be annihilated by a simple Givens rotation
and the resulting upper bidiagonal system (the other subdiagonal elements having been removed in
previous iteration steps) can simply be solved, which leads to the MINRES method (see Paige and
Saunders [167]).
Another possibility is to solve the system Tiy = r(0)
2e(1)
, as in the CG method (Ti is
the upper i × i part of ¯Ti). Other than in CG we cannot rely on the existence of a Choleski
decomposition (since A is not positive definite). An alternative is then to decompose Ti by an
LQ-decomposition. This again leads to simple recurrences and the resulting method is known as
SYMMLQ (see Paige and Saunders [167]).
2.3.3 CG on the Normal Equations, CGNE and CGNR
The CGNE and CGNR methods are the simplest methods for nonsymmetric or indefinite systems.
Since other methods for such systems are in general rather more complicated than the Conjugate
Gradient method, transforming the system to a symmetric definite one and then applying the Con-
jugate Gradient method is attractive for its coding simplicity.
Theory
If a system of linear equations Ax = b has a nonsymmetric, possibly indefinite (but nonsingular),
coefficient matrix, one obvious attempt at a solution is to apply Conjugate Gradient to a related
symmetric positive definite system, AT
Ax = AT
b. While this approach is easy to understand
and code, the convergence speed of the Conjugate Gradient method now depends on the square of
the condition number of the original coefficient matrix. Thus the rate of convergence of the CG
procedure on the normal equations may be slow.
Several proposals have been made to improve the numerical stability of this method. The best
known is by Paige and Saunders [168] and is based upon applying the Lanczos method to the
auxiliary system
I A
AT
0
r
x
=
b
0
.
A clever execution of this scheme delivers the factors L and U of the LU-decomposition of the
tridiagonal matrix that would have been computed by carrying out the Lanczos procedure with
AT
A.
Another means for improving the numerical stability of this normal equations approach is sug-
gested by Bj¨orck and Elfving in [34]. The observation that the matrix AT
A is used in the construc-
tion of the iteration coefficients through an inner product like (p, AT
Ap) leads to the suggestion
that such an inner product be replaced by (Ap, Ap).](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-26-320.jpg)
![2.3. NONSTATIONARY ITERATIVE METHODS 17
2.3.4 Generalized Minimal Residual (GMRES)
The Generalized Minimal Residual method is an extension of MINRES (which is only applicable to
symmetric systems) to unsymmetric systems. Like MINRES, it generates a sequence of orthogonal
vectors, but in the absence of symmetry this can no longer be done with short recurrences; instead,
all previously computed vectors in the orthogonal sequence have to be retained. For this reason,
“restarted” versions of the method are used.
In the Conjugate Gradient method, the residuals form an orthogonal basis for the space
span{r(0)
, Ar(0)
, A2
r(0)
, . . .}. In GMRES, this basis is formed explicitly:
w(i)
= Av(i)
for k = 1, . . . , i
w(i)
= w(i)
− (w(i)
, v(k)
)v(k)
end
v(i+1)
= w(i)
/ w(i)
The reader may recognize this as a modified Gram-Schmidt orthogonalization. Applied to the
Krylov sequence {Ak
r(0)
} this orthogonalization is called the “Arnoldi method” [6]. The inner
product coefficients (w(i)
, v(k)
) and w(i)
are stored in an upper Hessenberg matrix.
The GMRES iterates are constructed as
x(i)
= x(0)
+ y1v(1)
+ · · · + yiv(i)
,
where the coefficients yk have been chosen to minimize the residual norm b − Ax(i)
. The
GMRES algorithm has the property that this residual norm can be computed without the iterate
having been formed. Thus, the expensive action of forming the iterate can be postponed until the
residual norm is deemed small enough.
The pseudocode for the restarted GMRES(m) algorithm with preconditioner M is given in
Figure 2.6.
Theory
The Generalized Minimum Residual (GMRES) method is designed to solve nonsymmetric linear
systems (see Saad and Schultz [188]). The most popular form of GMRES is based on the modified
Gram-Schmidt procedure, and uses restarts to control storage requirements.
If no restarts are used, GMRES (like any orthogonalizing Krylov-subspace method) will con-
verge in no more than n steps (assuming exact arithmetic). Of course this is of no practical value
when n is large; moreover, the storage and computational requirements in the absence of restarts
are prohibitive. Indeed, the crucial element for successful application of GMRES(m) revolves
around the decision of when to restart; that is, the choice of m. Unfortunately, there exist examples
for which the method stagnates and convergence takes place only at the nth step. For such systems,
any choice of m less than n fails to converge.
Saad and Schultz [188] have proven several useful results. In particular, they show that if the
coefficient matrix A is real and nearly positive definite, then a “reasonable” value for m may be
selected. Implications of the choice of m are discussed below.
Implementation
A common implementation of GMRES is suggested by Saad and Schultz in [188] and relies on
using modified Gram-Schmidt orthogonalization. Householder transformations, which are rela-
tively costly but stable, have also been proposed. The Householder approach results in a three-fold
increase in work; however, convergence may be better, especially for ill-conditioned systems (see
Walker [212]). From the point of view of parallelism, Gram-Schmidt orthogonalization may be](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-27-320.jpg)
![18 CHAPTER 2. ITERATIVE METHODS
x(0)
is an initial guess
for j = 1, 2, ....
Solve r from Mr = b − Ax(0)
v(1)
= r/ r 2
s := r 2e1
for i = 1, 2, ..., m
Solve w from Mw = Av(i)
for k = 1, ..., i
hk,i = (w, v(k)
)
w = w − hk,iv(k)
end
hi+1,i = w 2
v(i+1)
= w/hi+1,i
apply J1, ..., Ji−1 on (h1,i, ..., hi+1,i)
construct Ji, acting on ith and (i + 1)st component
of h.,i, such that (i + 1)st component of Jih.,i is 0
s := Jis
if s(i + 1) is small enough then (UPDATE(˜x, i) and quit)
end
UPDATE(˜x, m)
end
In this scheme UPDATE(˜x, i)
replaces the following computations:
Compute y as the solution of Hy = ˜s, in which
the upper i × i triangular part of H has hi,j as
its elements (in least squares sense if H is singular),
˜s represents the first i components of s
˜x = x(0)
+ y1v(1)
+ y2v(2)
+ ... + yiv(i)
s(i+1)
= b − A˜x 2
if ˜x is an accurate enough approximation then quit
else x(0)
= ˜x
Figure 2.6: The Preconditioned GMRES(m) Method
preferred, giving up some stability for better parallelization properties (see Demmel, Heath and
Van der Vorst [66]). Here we adopt the Modified Gram-Schmidt approach.
The major drawback to GMRES is that the amount of work and storage required per iteration
rises linearly with the iteration count. Unless one is fortunate enough to obtain extremely fast con-
vergence, the cost will rapidly become prohibitive. The usual way to overcome this limitation is by
restarting the iteration. After a chosen number (m) of iterations, the accumulated data are cleared
and the intermediate results are used as the initial data for the next m iterations. This procedure is
repeated until convergence is achieved. The difficulty is in choosing an appropriate value for m. If
m is “too small”, GMRES(m) may be slow to converge, or fail to converge entirely. A value of m
that is larger than necessary involves excessive work (and uses more storage). Unfortunately, there
are no definite rules governing the choice of m—choosing when to restart is a matter of experience.
For a discussion of GMRES for vector and shared memory computers see Dongarra et al. [70];](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-28-320.jpg)
![2.3. NONSTATIONARY ITERATIVE METHODS 19
for more general architectures, see Demmel, Heath and Van der Vorst [66].
2.3.5 BiConjugate Gradient (BiCG)
The Conjugate Gradient method is not suitable for nonsymmetric systems because the residual
vectors cannot be made orthogonal with short recurrences (for proof of this see Voevodin [211]
or Faber and Manteuffel [95]). The GMRES method retains orthogonality of the residuals by
using long recurrences, at the cost of a larger storage demand. The BiConjugate Gradient method
takes another approach, replacing the orthogonal sequence of residuals by two mutually orthogonal
sequences, at the price of no longer providing a minimization.
The update relations for residuals in the Conjugate Gradient method are augmented in the
BiConjugate Gradient method by relations that are similar but based on AT
instead of A. Thus
we update two sequences of residuals
r(i)
= r(i−1)
− αiAp(i)
, ˜r(i)
= ˜r(i−1)
− αiAT
˜p(i)
,
and two sequences of search directions
p(i)
= r(i−1)
+ βi−1p(i−1)
, ˜p(i)
= ˜r(i−1)
+ βi−1 ˜p(i−1)
.
The choices
αi =
˜r(i−1)T
r(i−1)
˜p(i)T
Ap(i)
, βi =
˜r(i)T
r(i)
˜r(i−1)T
r(i−1)
ensure the bi-orthogonality relations
˜r(i)T
r(j)
= ˜p(i)T
Ap(j)
= 0 if i = j.
The pseudocode for the Preconditioned BiConjugate Gradient Method with preconditioner M
is given in Figure 2.7.
Convergence
Few theoretical results are known about the convergence of BiCG. For symmetric positive defi-
nite systems the method delivers the same results as CG, but at twice the cost per iteration. For
nonsymmetric matrices it has been shown that in phases of the process where there is significant
reduction of the norm of the residual, the method is more or less comparable to full GMRES (in
terms of numbers of iterations) (see Freund and Nachtigal [101]). In practice this is often con-
firmed, but it is also observed that the convergence behavior may be quite irregular, and the method
may even break down. The breakdown situation due to the possible event that z(i−1)T
˜r(i−1)
≈ 0
can be circumvented by so-called look-ahead strategies (see Parlett, Taylor and Liu [171]). This
leads to complicated codes and is beyond the scope of this book. The other breakdown situation,
˜p(i)T
q(i)
≈ 0, occurs when the LU-decomposition fails (see the theory subsection of §2.3.1), and
can be repaired by using another decomposition. This is done in QMR (see §2.3.6).
Sometimes, breakdown or near-breakdown situations can be satisfactorily avoided by a restart
at the iteration step immediately before the (near-) breakdown step. Another possibility is to switch
to a more robust (but possibly more expensive) method, like GMRES.
Implementation
BiCG requires computing a matrix-vector product Ap(k)
and a transpose product AT
˜p(k)
. In some
applications the latter product may be impossible to perform, for instance if the matrix is not formed
explicitly and the regular product is only given in operation form, for instance as a function call
evaluation.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-29-320.jpg)
![20 CHAPTER 2. ITERATIVE METHODS
Compute r(0)
= b − Ax(0)
for some initial guess x(0)
.
Choose ˜r(0)
(for example, ˜r(0)
= r(0)
).
for i = 1, 2, . . .
solve Mz(i−1)
= r(i−1)
solve MT
˜z(i−1)
= ˜r(i−1)
ρi−1 = z(i−1)T
˜r(i−1)
if ρi−1 = 0, method fails
if i = 1
p(i)
= z(i−1)
˜p(i)
= ˜z(i−1)
else
βi−1 = ρi−1/ρi−2
p(i)
= z(i−1)
+ βi−1p(i−1)
˜p(i)
= ˜z(i−1)
+ βi−1 ˜p(i−1)
endif
q(i)
= Ap(i)
˜q(i)
= AT
˜p(i)
αi = ρi−1/˜p(i)T
q(i)
x(i)
= x(i−1)
+ αip(i)
r(i)
= r(i−1)
− αiq(i)
˜r(i)
= ˜r(i−1)
− αi ˜q(i)
check convergence; continue if necessary
end
Figure 2.7: The Preconditioned BiConjugate Gradient Method
In a parallel environment, the two matrix-vector products can theoretically be performed si-
multaneously; however, in a distributed-memory environment, there will be extra communication
costs associated with one of the two matrix-vector products, depending upon the storage scheme
for A. A duplicate copy of the matrix will alleviate this problem, at the cost of doubling the storage
requirements for the matrix.
Care must also be exercised in choosing the preconditioner, since similar problems arise during
the two solves involving the preconditioning matrix.
It is difficult to make a fair comparison between GMRES and BiCG. GMRES really minimizes
a residual, but at the cost of increasing work for keeping all residuals orthogonal and increasing
demands for memory space. BiCG does not minimize a residual, but often its accuracy is com-
parable to GMRES, at the cost of twice the amount of matrix vector products per iteration step.
However, the generation of the basis vectors is relatively cheap and the memory requirements are
modest. Several variants of BiCG have been proposed that increase the effectiveness of this class
of methods in certain circumstances. These variants (CGS and Bi-CGSTAB) will be discussed in
coming subsections.
2.3.6 Quasi-Minimal Residual (QMR)
The BiConjugate Gradient method often displays rather irregular convergence behavior. Moreover,
the implicit LU decomposition of the reduced tridiagonal system may not exist, resulting in break-
down of the algorithm. A related algorithm, the Quasi-Minimal Residual method of Freund and
Nachtigal [101], [102] attempts to overcome these problems. The main idea behind this algorithm](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-30-320.jpg)
![2.3. NONSTATIONARY ITERATIVE METHODS 21
is to solve the reduced tridiagonal system in a least squares sense, similar to the approach fol-
lowed in GMRES. Since the constructed basis for the Krylov subspace is bi-orthogonal, rather than
orthogonal as in GMRES, the obtained solution is viewed as a quasi-minimal residual solution,
which explains the name. Additionally, QMR uses look-ahead techniques to avoid breakdowns in
the underlying Lanczos process, which makes it more robust than BiCG.
Convergence
The convergence behavior of QMR is typically much smoother than for BiCG. Freund and Nachti-
gal [101] present quite general error bounds which show that QMR may be expected to converge
about as fast as GMRES. From a relation between the residuals in BiCG and QMR (Freund and
Nachtigal [101, relation (5.10)]) one may deduce that at phases in the iteration process where BiCG
makes significant progress, QMR has arrived at about the same approximation for ˆx. On the other
hand, when BiCG makes no progress at all, QMR may still show slow convergence.
The look-ahead steps in the QMR method prevent breakdown in all cases but the so-called
“incurable breakdown”, where no number of look-ahead steps would yield a next iterate.
Implementation
The pseudocode for the Preconditioned Quasi Minimal Residual Method, with preconditioner
M = M1M2, is given in Figure 2.8. This algorithm follows the two term recurrence version
without look-ahead, presented by Freund and Nachtigal [102] as Algorithm 7.1. This version of
QMR is simpler to implement than the full QMR method with look-ahead, but it is susceptible
to breakdown of the underlying Lanczos process. (Other implementational variations are whether
to scale Lanczos vectors or not, or to use three-term recurrences instead of coupled two-term re-
currences. Such decisions usually have implications for the stability and the efficiency of the algo-
rithm.) A professional implementation of QMR with look-ahead is given in Freund and Nachtigal’s
QMRPACK, which is available through netlib; see Appendix A.
We have modified the algorithm to include a relatively inexpensive recurrence relation for the
computation of the residual vector. This requires a few extra vectors of storage and vector update
operations per iteration, but it avoids expending a matrix-vector product on the residual calculation.
Also, the algorithm has been modified so that only two full preconditioning steps are required
instead of three.
Computation of the residual is done for the convergence test. If one uses right (or post) precon-
ditioning, that is M1 = I, then a cheap upper bound for r(i)
can be computed in each iteration,
avoiding the recursions for r(i)
. For details, see Freund and Nachtigal [101, proposition 4.1]. This
upper bound may be pessimistic by a factor of at most
√
i + 1.
QMR has roughly the same problems with respect to vector and parallel implementation as
BiCG. The scalar overhead per iteration is slightly more than for BiCG. In all cases where the
slightly cheaper BiCG method converges irregularly (but fast enough), there seems little reason to
avoid QMR.
2.3.7 Conjugate Gradient Squared Method (CGS)
In BiCG, the residual vector r(i)
can be regarded as the product of r(0)
and an ith degree polynomial
in A, that is
r(i)
= Pi(A)r(0)
. (2.13)
This same polynomial satisfies ˜r(i)
= Pi(A)˜r(0)
so that
ρi = (˜r(i)
, r(i)
) = (Pi(AT
)˜r(0)
, Pi(A)r(0)
) = (˜r(i)
, P2
i (A)r(0)
). (2.14)](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-31-320.jpg)
![2.3. NONSTATIONARY ITERATIVE METHODS 23
Compute r(0)
= b − Ax(0)
for some initial guess x(0)
Choose ˜r (for example, ˜r = r(0)
)
for i = 1, 2, . . .
ρi−1 = ˜rT
r(i−1)
if ρi−1 = 0 method fails
if i = 1
u(1)
= r(0)
p(1)
= u(1)
else
βi−1 = ρi−1/ρi−2
u(i)
= r(i−1)
+ βi−1q(i−1)
p(i)
= u(i)
+ βi−1(q(i−1)
+ βi−1p(i−1)
)
endif
solve M ˆp = p(i)
ˆv = Aˆp
αi = ρi−1/˜rT
ˆv
q(i)
= u(i)
− αiˆv
solve M ˆu = u(i)
+ q(i)
x(i)
= x(i−1)
+ αi ˆu
ˆq = Aˆu
r(i)
= r(i−1)
− αi ˆq
check convergence; continue if necessary
end
Figure 2.9: The Preconditioned Conjugate Gradient Squared Method
This suggests that if Pi(A) reduces r(0)
to a smaller vector r(i)
, then it might be advantageous
to apply this “contraction” operator twice, and compute P2
i (A)r(0)
. Equation (2.14) shows that the
iteration coefficients can still be recovered from these vectors, and it turns out to be easy to find
the corresponding approximations for x. This approach leads to the Conjugate Gradient Squared
method (see Sonneveld [191]).
Convergence
Often one observes a speed of convergence for CGS that is about twice as fast as for BiCG, which is
in agreement with the observation that the same “contraction” operator is applied twice. However,
there is no reason that the “contraction” operator, even if it really reduces the initial residual r(0)
,
should also reduce the once reduced vector r(k)
= Pk(A)r(0)
. This is evidenced by the often highly
irregular convergence behavior of CGS. One should be aware of the fact that local corrections to
the current solution may be so large that cancellation effects occur. This may lead to a less accurate
solution than suggested by the updated residual (see Van der Vorst [205]). The method tends to
diverge if the starting guess is close to the solution.
Implementation
CGS requires about the same number of operations per iteration as BiCG, but does not involve
computations with AT
. Hence, in circumstances where computation with AT
is impractical, CGS
may be attractive.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-33-320.jpg)
![24 CHAPTER 2. ITERATIVE METHODS
Compute r(0)
= b − Ax(0)
for some initial guess x(0)
Choose ˜r (for example, ˜r = r(0)
)
for i = 1, 2, . . .
ρi−1 = ˜rT
r(i−1)
if ρi−1 = 0 method fails
if i = 1
p(i)
= r(i−1)
else
βi−1 = (ρi−1/ρi−2)(αi−1/ωi−1)
p(i)
= r(i−1)
+ βi−1(p(i−1)
− ωi−1v(i−1)
)
endif
solve M ˆp = p(i)
v(i)
= Aˆp
αi = ρi−1/˜rT
v(i)
s = r(i−1)
− αiv(i)
check norm of s; if small enough: set x(i)
= x(i−1)
+ αi ˆp and stop
solve Mˆs = s
t = Aˆs
ωi = tT
s/tT
t
x(i)
= x(i−1)
+ αi ˆp + ωiˆs
r(i)
= s − ωit
check convergence; continue if necessary
for continuation it is necessary that ωi = 0
end
Figure 2.10: The Preconditioned BiConjugate Gradient Stabilized Method
The pseudocode for the Preconditioned Conjugate Gradient Squared Method with precondi-
tioner M is given in Figure 2.9.
2.3.8 BiConjugate Gradient Stabilized (Bi-CGSTAB)
The BiConjugate Gradient Stabilized method (Bi-CGSTAB) was developed to solve nonsymmet-
ric linear systems while avoiding the often irregular convergence patterns of the Conjugate Gra-
dient Squared method (see Van der Vorst [205]). Instead of computing the CGS sequence i →
P2
i (A)r(0)
, Bi-CGSTAB computes i → Qi(A)Pi(A)r(0)
where Qi is an ith degree polynomial
describing a steepest descent update.
Convergence
Bi-CGSTAB often converges about as fast as CGS, sometimes faster and sometimes not. CGS can
be viewed as a method in which the BiCG “contraction” operator is applied twice. Bi-CGSTAB
can be interpreted as the product of BiCG and repeatedly applied GMRES(1). At least locally, a
residual vector is minimized, which leads to a considerably smoother convergence behavior. On
the other hand, if the local GMRES(1) step stagnates, then the Krylov subspace is not expanded,
and Bi-CGSTAB will break down. This is a breakdown situation that can occur in addition to the
other breakdown possiblities in the underlying BiCG algorithm. This type of breakdown may be
avoided by combining BiCG with other methods, i.e., by selecting other values for ωi (see the](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-34-320.jpg)
![2.3. NONSTATIONARY ITERATIVE METHODS 25
algorithm). One such alternative is Bi-CGSTAB2 (see Gutknecht [114]); more general approaches
are suggested by Sleijpen and Fokkema in [189].
Note that Bi-CGSTAB has two stopping tests: if the method has already converged at the first
test on the norm of s, the subsequent update would be numerically questionable. Additionally,
stopping on the first test saves a few unnecessary operations, but this is of minor importance.
Implementation
Bi-CGSTAB requires two matrix-vector products and four inner products, i.e., two inner products
more than BiCG and CGS.
The pseudocode for the Preconditioned BiConjugate Gradient Stabilized Method with precon-
ditioner M is given in Figure 2.10.
2.3.9 Chebyshev Iteration
Chebyshev Iteration is another method for solving nonsymmetric problems (see Golub and
Van Loan [108, §10.1.5] and Varga [209, Chapter 5]). Chebyshev Iteration avoids the computation
of inner products as is necessary for the other nonstationary methods. For some distributed memory
architectures these inner products are a bottleneck with respect to efficiency. The price one pays
for avoiding inner products is that the method requires enough knowledge about the spectrum
of the coefficient matrix A that an ellipse enveloping the spectrum can be identified; however
this difficulty can be overcome via an adaptive construction developed by Manteuffel [145], and
implemented by Ashby [7]. Chebyshev iteration is suitable for any nonsymmetric linear system
for which the enveloping ellipse does not include the origin.
Comparison with other methods
Comparing the pseudocode for Chebyshev Iteration with the pseudocode for the Conjugate Gra-
dient method shows a high degree of similarity, except that no inner products are computed in
Chebyshev Iteration.
Scalars c and d must be selected so that they define a family of ellipses with common center
d > 0 and foci d + c and d − c which contain the ellipse that encloses the spectrum (or more
general, field of values) of A and for which the rate r of convergence is minimal:
r =
a +
√
a2 − c2
d +
√
d2 − c2
, (2.15)
where a is the length of the x-axis of the ellipse.
We provide code in which it is assumed that c and d are known. For code including the adap-
tive detemination of these iteration parameters the reader is referred to Ashby [7]. The Chebyshev
method has the advantage over GMRES that only short recurrences are used. On the other hand,
GMRES is guaranteed to generate the smallest residual over the current search space. The BiCG
methods, which also use short recurrences, do not minimize the residual in a suitable norm; how-
ever, unlike Chebyshev iteration, they do not require estimation of parameters (the spectrum of A).
Finally, GMRES and BiCG may be more effective in practice, because of superlinear convergence
behavior, which cannot be expected for Chebyshev.
For symmetric positive definite systems the “ellipse” enveloping the spectrum degenerates to
the interval [λmin, λmax] on the positive x-axis, where λmin and λmax are the smallest and largest
eigenvalues of M−1
A. In circumstances where the computation of inner products is a bottleneck,
it may be advantageous to start with CG, compute estimates of the extremal eigenvalues from the
CG coefficients, and then after sufficient convergence of these approximations switch to Chebyshev
Iteration. A similar strategy may be adopted for a switch from GMRES, or BiCG-type methods, to
Chebyshev Iteration.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-35-320.jpg)
![26 CHAPTER 2. ITERATIVE METHODS
Compute r(0)
= b − Ax(0)
for some initial guess x(0)
.
d = (λmax + λmin)/2, c = (λmax − λmin)/2.
for i = 1, 2, . . .
solve Mz(i−1)
= r(i)
.
if i = 1
p(1)
= z(0)
α1 = 2/d
else
βi−1 = (cαi−1/2)2
αi = 1/(d − βi−1)
p(i)
= z(i−1)
+ βi−1p(i−1)
.
endif
x(i)
= x(i−1)
+ αip(i)
.
r(i)
= b − Ax(i)
(= r(i−1)
− αiAp(i)
).
check convergence; continue if necessary
end
Figure 2.11: The Preconditioned Chebyshev Method
Convergence
In the symmetric case (where A and the preconditioner M are both symmetric) for the Chebyshev
Iteration we have the same upper bound as for the Conjugate Gradient method, provided c and d are
computed from λmin and λmax (the extremal eigenvalues of the preconditioned matrix M−1
A).
There is a severe penalty for overestimating or underestimating the field of values. For ex-
ample, if in the symmetric case λmax is underestimated, then the method may diverge; if it is
overestimated then the result may be very slow convergence. Similar statements can be made for
the nonsymmetric case. This implies that one needs fairly accurate bounds on the spectrum of
M−1
A for the method to be effective (in comparison with CG or GMRES).
Implementation
In Chebyshev Iteration the iteration parameters are known as soon as one knows the ellipse con-
taining the eigenvalues (or rather, the field of values) of the operator. Therefore the computation
of inner products, as is necessary in methods like GMRES or CG, is avoided. This avoids the
synchronization points required of CG-type methods, so machines with hierarchical or distributed
memory may achieve higher performance (it also suggests strong parallelization properties; for a
discussion of this see Saad [184], and Dongarra, et al. [70]). Specifically, as soon as some segment
of w is computed, we may begin computing, in sequence, corresponding segments of p, x, and r.
The pseudocode for the Preconditioned Chebyshev Method with preconditioner M is given
in Figure 2.11. It handles the case of a symmetric positive definite coefficient matrix A. The
eigenvalues of M−1
A are assumed to be all real and in the interval [λmin, λmax], which does not
include zero.
2.4 Computational Aspects of the Methods
Efficient solution of a linear system is largely a function of the proper choice of iterative method.
However, to obtain good performance, consideration must also be given to the computational ker-](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-36-320.jpg)
![2.4. COMPUTATIONAL ASPECTS OF THE METHODS 27
Matrix-
Inner Vector Precond
Method Product SAXPY Product Solve
JACOBI 1a
GS 1 1a
SOR 1 1a
CG 2 3 1 1
GMRES i + 1 i + 1 1 1
BiCG 2 5 1/1 1/1
QMR 2 8+4bc
1/1 1/1
CGS 2 6 2 2
Bi-CGSTAB 4 6 2 2
CHEBYSHEV 2 1 1
Table 2.1: Summary of Operations for Iteration i. “a/b” means “a” multiplications with the matrix
and “b” with its transpose.
aThis method performs no real matrix vector product or preconditioner solve, but the number of operations is equivalent
to a matrix-vector multiply.
bTrue SAXPY operations + vector scalings.
cLess for implementations that do not recursively update the residual.
nels of the method and how efficiently they can be executed on the target architecture. This point
is of particular importance on parallel architectures; see §4.4.
Iterative methods are very different from direct methods in this respect. The performance of
direct methods, both for dense and sparse systems, is largely that of the factorization of the matrix.
This operation is absent in iterative methods (although preconditioners may require a setup phase),
and with it, iterative methods lack dense matrix suboperations. Since such operations can be exe-
cuted at very high efficiency on most current computer architectures, we expect a lower flop rate for
iterative than for direct methods. (Dongarra and Van der Vorst [73] give some experimental results
about this, and provide a benchmark code for iterative solvers.) Furthermore, the basic operations
in iterative methods often use indirect addressing, depending on the data structure. Such operations
also have a relatively low efficiency of execution.
However, this lower efficiency of execution does not imply anything about the total solution
time for a given system. Furthermore, iterative methods are usually simpler to implement than
direct methods, and since no full factorization has to be stored, they can handle much larger sytems
than direct methods.
In this section we summarize for each method
• Matrix properties. Not every method will work on every problem type, so knowledge of
matrix properties is the main criterion for selecting an iterative method.
• Computational kernels. Different methods involve different kernels, and depending on the
problem or target computer architecture this may rule out certain methods.
Table 2.2 lists the storage required for each method (without preconditioning). Note that we
are not including the original system Ax = b and we ignore scalar storage.
1. Jacobi Method
• Extremely easy to use, but unless the matrix is “strongly” diagonally dominant, this
method is probably best only considered as an introduction to iterative methods or as a
preconditioner in a nonstationary method.
• Trivial to parallelize.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-37-320.jpg)
![30 CHAPTER 2. ITERATIVE METHODS
• This method requires some explicit knowledge of the spectrum (or field of values); in
the symmetric case the iteration parameters are easily obtained from the two extremal
eigenvalues, which can be estimated either directly from the matrix, or from applying
a few iterations of the Conjugate Gradient Method.
• The computational structure is similar to that of CG, but there are no synchronization
points.
• The Adaptive Chebyshev method can be used in combination with methods as CG or
GMRES, to continue the iteration once suitable bounds on the spectrum have been
obtained from these methods.
Selecting the “best” method for a given class of problems is largely a matter of trial and error.
It also depends on how much storage one has available (GMRES), on the availability of AT
(BiCG
and QMR), and on how expensive the matrix vector products (and Solve steps with M) are in
comparison to SAXPYs and inner products. If these matrix vector products are relatively expensive,
and if sufficient storage is available then it may be attractive to use GMRES and delay restarting as
much as possible.
Table 2.1 shows the type of operations performed per iteration. Based on the particular prob-
lem or data structure, the user may observe that a particular operation could be performed more
efficiently.
2.5 A short history of Krylov methods1
Methods based on orthogonalization were developed by a number of authors in the early ’50s.
Lanczos’ method [141] was based on two mutually orthogonal vector sequences, and his motivation
came from eigenvalue problems. In that context, the most prominent feature of the method is that it
reduces the original matrix to tridiagonal form. Lanczos later applied his method to solving linear
systems, in particular symmetric ones [142]. An important property for proving convergence of
the method when solving linear systems is that the iterates are related to the initial residual by
multiplication with a polynomial in the coefficient matrix.
The joint paper by Hestenes and Stiefel [121], after their independent discovery of the same
method, is the classical description of the conjugate gradient method for solving linear systems.
Although error-reduction properties are proved, and experiments showing premature convergence
are reported, the conjugate gradient method is presented here as a direct method, rather than an
iterative method.
This Hestenes/Stiefel method is closely related to a reduction of the Lanczos method to sym-
metric matrices, reducing the two mutually orthogonal sequences to one orthogonal sequence, but
there is an important algorithmic difference. Whereas Lanczos used three-term recurrences, the
method by Hestenes and Stiefel uses coupled two-term recurrences. By combining the two two-
term recurrences (eliminating the “search directions”) the Lanczos method is obtained.
A paper by Arnoldi [6] further discusses the Lanczos biorthogonalization method, but it also
presents a new method, combining features of the Lanczos and Hestenes/Stiefel methods. Like
the Lanczos method it is applied to nonsymmetric systems, and it does not use search directions.
Like the Hestenes/Stiefel method, it generates only one, self-orthogonal sequence. This last fact,
combined with the asymmetry of the coefficient matrix means that the method no longer effects a
reduction to tridiagonal form, but instead one to upper Hessenberg form. Presented as “minimized
iterations in the Galerkin method” this algorithm has become known as the Arnoldi algorithm.
The conjugate gradient method received little attention as a practical method for some time,
partly because of a misperceived importance of the finite termination property. Reid [178] pointed
1For a more detailed account of the early history of CG methods, we refer the reader to Golub and O’Leary [107] and
Hestenes [122].](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-40-320.jpg)
![2.6. SURVEY OF RECENT KRYLOV METHODS 31
out that the most important application area lay in sparse definite systems, and this renewed the
interest in the method.
Several methods have been developed in later years that employ, most often implicitly, the
upper Hessenberg matrix of the Arnoldi method. For an overview and characterization of these
orthogonal projection methods for nonsymmetric systems see Ashby, Manteuffel and Saylor [10],
Saad and Schultz [187], and Jea and Young [124].
Fletcher [97] proposed an implementation of the Lanczos method, similar to the Conjugate Gra-
dient method, with two coupled two-term recurrences, which he named the bi-conjugate gradient
method (BiCG).
2.6 Survey of recent Krylov methods
Research into the design of Krylov subspace methods for solving nonsymmetric linear systems is
an active field of research and new methods are still emerging. In this book, we have included only
the best known and most popular methods, and in particular those for which extensive computa-
tional experience has been gathered. In this section, we shall briefly highlight some of the recent
developments and other methods not treated here. A survey of methods up to about 1991 can be
found in Freund, Golub and Nachtigal [105]. Two more recent reports by Meier-Yang [150] and
Tong [195] have extensive numerical comparisons among various methods, including several more
recent ones that have not been discussed in detail in this book.
Several suggestions have been made to reduce the increase in memory and computational costs
in GMRES. An obvious one is to restart (this one is included in §2.3.4): GMRES(m). Another
approach is to restrict the GMRES search to a suitable subspace of some higher-dimensional Krylov
subspace. Methods based on this idea can be viewed as preconditioned GMRES methods. The
simplest ones exploit a fixed polynomial preconditioner (see Johnson, Micchelli and Paul [125],
Saad [182], and Nachtigal, Reichel and Trefethen [158]). In more sophisticated approaches, the
polynomial preconditioner is adapted to the iterations (Saad [186]), or the preconditioner may
even be some other (iterative) method of choice (Van der Vorst and Vuik [207], Axelsson and
Vassilevski [24]). Stagnation is prevented in the GMRESR method (Van der Vorst and Vuik [207])
by including LSQR steps in some phases of the process. In De Sturler and Fokkema [63], part of
the optimality of GMRES is maintained in the hybrid method GCRO, in which the iterations of
the preconditioning method are kept orthogonal to the iterations of the underlying GCR method.
All these approaches have advantages for some problems, but it is far from clear a priori which
strategy is preferable in any given case.
Recent work has focused on endowing the BiCG method with several desirable properties: (1)
avoiding breakdown; (2) avoiding use of the transpose; (3) efficient use of matrix-vector products;
(4) smooth convergence; and (5) exploiting the work expended in forming the Krylov space with
AT
for further reduction of the residual.
As discussed before, the BiCG method can have two kinds of breakdown: Lanczos breakdown
(the underlying Lanczos process breaks down), and pivot breakdown (the tridiagonal matrix T
implicitly generated in the underlying Lanczos process encounters a zero pivot when Gaussian
elimination without pivoting is used to factor it). Although such exact breakdowns are very rare in
practice, near breakdowns can cause severe numerical stability problems.
The pivot breakdown is the easier one to overcome and there have been several approaches
proposed in the literature. It should be noted that for symmetric matrices, Lanczos breakdown
cannot occur and the only possible breakdown is pivot breakdown. The SYMMLQ and QMR
methods discussed in this book circumvent pivot breakdown by solving least squares systems.
Other methods tackling this problem can be found in Fletcher [97], Saad [180], Gutknecht [113],
and Bank and Chan [29, 28].
Lanczos breakdown is much more difficult to eliminate. Recently, considerable attention
has been given to analyzing the nature of the Lanczos breakdown (see Parlett [171], and](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-41-320.jpg)
![32 CHAPTER 2. ITERATIVE METHODS
Gutknecht [112, 115]), as well as various look-ahead techniques for remedying it (see Brezinski
and Sadok [39], Brezinski, Zaglia and Sadok [40, 41], Freund and Nachtigal [101], Parlett [171],
Nachtigal [159], Freund, Gutknecht and Nachtigal [100], Joubert [128], Freund, Golub and
Nachtigal [105], and Gutknecht [112, 115]). However, the resulting algorithms are usually too
complicated to give in template form (some codes of Freund and Nachtigal are available on
netlib.) Moreover, it is still not possible to eliminate breakdowns that require look-ahead
steps of arbitrary size (incurable breakdowns). So far, these methods have not yet received much
practical use but some form of look-ahead may prove to be a crucial component in future methods.
In the BiCG method, the need for matrix-vector multiplies with AT
can be inconvenient as
well as doubling the number of matrix-vector multiplies compared with CG for each increase in
the degree of the underlying Krylov subspace. Several recent methods have been proposed to over-
come this drawback. The most notable of these is the ingenious CGS method by Sonneveld [191]
discussed earlier, which computes the square of the BiCG polynomial without requiring AT
– thus
obviating the need for AT
. When BiCG converges, CGS is often an attractive, faster converging
alternative. However, CGS also inherits (and often magnifies) the breakdown conditions and the
irregular convergence of BiCG (see Van der Vorst [205]).
CGS also generated interest in the possibility of product methods, which generate iterates
corresponding to a product of the BiCG polynomial with another polynomial of the same de-
gree, chosen to have certain desirable properties but computable without recourse to AT
. The
Bi-CGSTAB method of Van der Vorst [205] is such an example, in which the auxiliary polynomial
is defined by a local minimization chosen to smooth the convergence behavior. Gutknecht [114]
noted that Bi-CGSTAB could be viewed as a product of BiCG and GMRES(1), and he suggested
combining BiCG with GMRES(2) for the even numbered iteration steps. This was anticipated to
lead to better convergence for the case where the eigenvalues of A are complex. A more efficient
and more robust variant of this approach has been suggested by Sleijpen and Fokkema in [189],
where they describe how to easily combine BiCG with any GMRES(m), for modest m.
Many other basic methods can also be squared. For example, by squaring the Lanczos pro-
cedure, Chan, de Pillis and Van der Vorst [45] obtained transpose-free implementations of BiCG
and QMR. By squaring the QMR method, Freund and Szeto [103] derived a transpose-free QMR
squared method which is quite competitive with CGS but with much smoother convergence. Un-
fortunately, these methods require an extra matrix-vector product per step (three instead of two)
which makes them less efficient.
In addition to Bi-CGSTAB, several recent product methods have been designed to smooth the
convergence of CGS. One idea is to use the quasi-minimal residual (QMR) principle to obtain
smoothed iterates from the Krylov subspace generated by other product methods. Freund [104]
proposed such a QMR version of CGS, which he called TFQMR. Numerical experiments show
that TFQMR in most cases retains the desirable convergence features of CGS while correcting its
erratic behavior. The transpose free nature of TFQMR, its low computational cost and its smooth
convergence behavior make it an attractive alternative to CGS. On the other hand, since the BiCG
polynomial is still used, TFQMR breaks down whenever CGS does. One possible remedy would be
to combine TFQMR with a look-ahead Lanczos technique but this appears to be quite complicated
and no methods of this kind have yet appeared in the literature. Recently, Chan et. al. [46] derived
a similar QMR version of Van der Vorst’s Bi-CGSTAB method, which is called QMRCGSTAB.
These methods offer smoother convergence over CGS and Bi-CGSTAB with little additional cost.
There is no clear best Krylov subspace method at this time, and there will never be a best
overall Krylov subspace method. Each of the methods is a winner in a specific problem class, and
the main problem is to identify these classes and to construct new methods for uncovered classes.
The paper by Nachtigal, Reddy and Trefethen [157] shows that for any of a group of methods (CG,
BiCG, GMRES, CGNE, and CGS), there is a class of problems for which a given method is the
winner and another one is the loser. This shows clearly that there will be no ultimate method. The
best we can hope for is some expert system that guides the user in his/her choice. Hence, iterative
methods will never reach the robustness of direct methods, nor will they beat direct methods for](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-42-320.jpg)
![36 CHAPTER 3. PRECONDITIONERS
3.1.2 Left and right preconditioning
The above transformation of the linear system A → M−1
A is often not what is used in practice.
For instance, the matrix M−1
A is not symmetric, so, even if A and M are, the conjugate gradients
method is not immediately applicable to this system. The method as described in figure 2.5 reme-
dies this by employing the M−1
-inner product for orthogonalization of the residuals. The theory
of the cg method is then applicable again.
All cg-type methods in this book, with the exception of QMR, have been derived with such a
combination of preconditioned iteration matrix and correspondingly changed inner product.
Another way of deriving the preconditioned conjugate gradients method would be to split the
preconditioner as M = M1M2 and to transform the system as
M−1
1 AM−1
2 (M2x) = M−1
1 b.
If M is symmetric and M1 = Mt
2, it is obvious that we now have a method with a symmetric
iteration matrix, hence the conjugate gradients method can be applied.
Remarkably, the splitting of M is in practice not needed. By rewriting the steps of the method
(see for instance Axelsson and Barker [14, pgs. 16,29] or Golub and Van Loan [108, §10.3]) it is
usually possible to reintroduce a computational step
solve u from Mu = v,
that is, a step that applies the preconditioner in its entirety.
There is a different approach to preconditioning, which is much easier to derive. Consider again
the system.
M−1
1 AM−1
2 (M2x) = M−1
1 b.
The matrices M1 and M2 are called the left- and right preconditioners, respectively, and we can
simpy apply an unpreconditioned iterative method to this system. Only two additional actions
r0 ← M−1
1 r0 before the iterative process and xn ← M−1
2 xn after are necessary.
Thus we arrive at the following schematic for deriving a left/right preconditioned iterative
method from any of the symmetrically preconditioned methods in this book.
1. Take a preconditioned iterative method, and replace every occurence of M by I.
2. Remove any vectors from the algorithm that have become duplicates in the previous step.
3. Replace every occurrence of A in the method by M−1
1 AM−1
2 .
4. After the calculation of the initial residual, add the step
r0 ← M−1
1 r0.
5. At the end of the method, add the step
x ← M−1
2 x,
where x is the final calculated solution.
It should be noted that such methods cannot be made to reduce to the algorithms given in section 2.3
by such choices as M1 = Mt
2 or M1 = I.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-46-320.jpg)
![3.2. JACOBI PRECONDITIONING 37
3.2 Jacobi Preconditioning
The simplest preconditioner consists of just the diagonal of the matrix:
mi,j =
ai,i if i = j
0 otherwise.
This is known as the (point) Jacobi preconditioner.
It is possible to use this preconditioner without using any extra storage beyond that of the matrix
itself. However, division operations are usually quite costly, so in practice storage is allocated for
the reciprocals of the matrix diagonal. This strategy applies to many preconditioners below.
3.2.1 Block Jacobi Methods
Block versions of the Jacobi preconditioner can be derived by a partitioning of the variables. If the
index set S = {1, . . . , n} is partitioned as S = i Si with the sets Si mutually disjoint, then
mi,j =
ai,j if i and j are in the same index subset
0 otherwise.
The preconditioner is now a block-diagonal matrix.
Often, natural choices for the partitioning suggest themselves:
• In problems with multiple physical variables per node, blocks can be formed by grouping the
equations per node.
• In structured matrices, such as those from partial differential equations on regular grids, a
partitioning can be based on the physical domain. Examples are a partitioning along lines in
the 2D case, or planes in the 3D case. This will be discussed further in §3.4.3.
• On parallel computers it is natural to let the partitioning coincide with the division of vari-
ables over the processors.
3.2.2 Discussion
Jacobi preconditioners need very little storage, even in the block case, and they are easy to imple-
ment. Additionally, on parallel computers they don’t present any particular problems.
On the other hand, more sophisticated preconditioners usually yield a larger improvement.1
3.3 SSOR preconditioning
The SSOR preconditioner2
like the Jacobi preconditioner, can be derived from the coefficient ma-
trix without any work.
If the original, symmetric, matrix is decomposed as
A = D + L + LT
1Under certain conditions, one can show that the point Jacobi algorithm is optimal, or close to optimal, in the sense
of reducing the condition number, among all preconditioners of diagonal form. This was shown by Forsythe and Strauss
for matrices with Property A [98], and by van der Sluis [196] for general sparse matrices. For extensions to block Jacobi
preconditioners, see Demmel [65] and Elsner [94].
2The SOR and Gauss-Seidel matrices are never used as preconditioners, for a rather technical reason. SOR-
preconditioning with optimal ω maps the eigenvalues of the coefficient matrix to a circle in the complex plane; see Hageman
and Young [119, §9.3]. In this case no polynomial acceleration is possible, i.e., the accelerating polynomial reduces to the
trivial polynomial Pn(x) = xn, and the resulting method is simply the stationary SOR method. Recent research by Eier-
mann and Varga [83] has shown that polynomial acceleration of SOR with suboptimal ω will yield no improvement over
simple SOR with optimal ω.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-47-320.jpg)
![38 CHAPTER 3. PRECONDITIONERS
in its diagonal, lower, and upper triangular part, the SSOR matrix is defined as
M = (D + L)D−1
(D + L)T
,
or, parametrized by ω
M(ω) =
1
2 − ω
(
1
ω
D + L)(
1
ω
D)−1
(
1
ω
D + L)T
.
The optimal value of the ω parameter, like the parameter in the SOR method, will reduce the
number of iterations to a lower order. Specifically, for second order eliptic problems a spectral
condition number κ(M−1
ωopt
A) = O( κ(A)) is attainable (see Axelsson and Barker [14, §1.4]).
In practice, however, the spectral information needed to calculate the optimal ω is prohibitively
expensive to compute.
The SSOR matrix is given in factored form, so this preconditioner shares many properties of
other factorization-based methods (see below). For instance, its suitability for vector processors or
parallel architectures depends strongly on the ordering of the variables. On the other hand, since
this factorization is given a priori, there is no possibility of breakdown as in the construction phase
of incomplete factorization methods.
3.4 Incomplete Factorization Preconditioners
A broad class of preconditioners is based on incomplete factorizations of the coefficient matrix.
We call a factorization incomplete if during the factorization process certain fill elements, nonzero
elements in the factorization in positions where the original matrix had a zero, have been ignored.
Such a preconditioner is then given in factored form M = LU with L lower and U upper triangular.
The efficacy of the preconditioner depends on how well M−1
approximates A−1
.
3.4.1 Creating an incomplete factorization
Incomplete factorizations are the first preconditioners we have encountered so far for which there
is a non-trivial creation stage. Incomplete factorizations may break down (attempted division by
zero pivot) or result in indefinite matrices (negative pivots) even if the full factorization of the same
matrix is guaranteed to exist and yield a positive definite matrix.
An incomplete factorization is guaranteed to exist for many factorization strategies if the orig-
inal matrix is an M-matrix. This was originally proved by Meijerink and Van der Vorst [151]; see
further Beauwens and Quenon [33], Manteuffel [146], and Van der Vorst [198].
In cases where pivots are zero or negative, strategies have been proposed such as substituting an
arbitrary positive number (see Kershaw [131]), or restarting the factorization on A + αI for some
positive value of α (see Manteuffel [146]).
An important consideration for incomplete factorization preconditioners is the cost of the fac-
torization process. Even if the incomplete factorization exists, the number of operations involved in
creating it is at least as much as for solving a system with such a coefficient matrix, so the cost may
equal that of one or more iterations of the iterative method. On parallel computers this problem is
aggravated by the generally poor parallel efficiency of the factorization.
Such factorization costs can be amortized if the iterative method takes many iterations, or if the
same preconditioner will be used for several linear systems, for instance in successive time steps
or Newton iterations.
Solving a system with an incomplete factorization preconditioner
Incomplete factorizations can be given in various forms. If M = LU (with L and U nonsingular
triangular matrices), solving a system proceeds in the usual way (figure 3.1), but often incomplete](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-48-320.jpg)
![40 CHAPTER 3. PRECONDITIONERS
Let S be the nonzero set {(i, j): aij = 0}
for i = 1, 2, . . .
set dii ← aii
for i = 1, 2, . . .
set dii ← 1/dii
for j = i + 1, i + 2, . . .
if (i, j) ∈ S and (j, i) ∈ S then
set djj ← djj − ajidiiaij
Figure 3.3: Construction of a D-ILU incomplete factorization preconditioner, storing the inverses
of the pivots
We can describe an incomplete factorization formally as
for each k, i, j > k: ai,j ←
ai,j − ai,ka−1
k,kak,j if (i, j) ∈ S
ai,j otherwise.
Meijerink and Van der Vorst [151] proved that, if A is an M-matrix, such a factorization exists for
any choice of S, and gives a symmetric positive definite matrix if A is symmetric positive definite.
Guidelines for allowing levels of fill were given by Meijerink and Van der Vorst in [152].
Fill-in strategies
There are two major strategies for accepting or discarding fill-in, one structural, and one numerical.
The structural strategy is that of accepting fill-in only to a certain level. As was already pointed out
above, any zero location (i, j) in A filling in (say in step k) is assigned a fill level value
fill(i, j) = 1 + max{fill(i, k), fill(k, i)}. (3.1)
If aij was already nonzero, the level value is not changed.
The numerical fill strategy is that of ‘drop tolerances’: fill is ignored if it is too small, for
a suitable definition of ‘small’. Although this definition makes more sense mathematically, it is
harder to implement in practice, since the amount of storage needed for the factorization is not
easy to predict. See [20, 156] for discussions of preconditioners using drop tolerances.
Simple cases: ILU(0) and D-ILU
For the ILU(0) method, the incomplete factorization produces no nonzero elements beyond the
sparsity structure of the original matrix, so that the preconditioner at worst takes exactly as much
space to store as the original matrix. In a simplified version of ILU(0), called D-ILU (Pom-
merell [173]), even less is needed. If not only we prohibit fill-in elements, but we also alter only
the diagonal elements (that is, any alterations of off-diagonal elements are ignored5
), we have the
following situation.
Splitting the coefficient matrix into its diagonal, lower triangular, and upper triangular parts as
A = DA+LA+UA, the preconditioner can be written as M = (D+LA)D−1
(D+UA) where D is
the diagonal matrix containing the pivots generated. Generating this preconditioner is described in
figure 3.3. Since we use the upper and lower triangle of the matrix unchanged, only storage space
for D is needed. In fact, in order to avoid division operations during the preconditioner solve stage
we store D−1
rather than D.
5In graph theoretical terms, ILU(0) and D-ILU coincide if the matrix graph contains no triangles.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-50-320.jpg)
![3.4. INCOMPLETE FACTORIZATION PRECONDITIONERS 41
Remark: the resulting lower and upper factors of the preconditioner have only nonzero elements
in the set S, but this fact is in general not true for the preconditioner M itself.
The fact that the D-ILU preconditioner contains the off-diagonal parts of the original matrix
was used by Eisenstat [90] to derive at a more efficient implementation of preconditioned CG.
This new implementation merges the application of the tridiagonal factors of the matrix and the
preconditioner, thereby saving a substantial number of operations per iteration.
Special cases: central differences
We will now consider the special case of a matrix derived from central differences on a Cartesian
product grid. In this case the ILU(0) and D-ILU factorizations coincide, and, as remarked above,
we only have to calculate the pivots of the factorization; other elements in the triangular factors are
equal to off-diagonal elements of A.
In the following we will assume a natural, line-by-line, ordering of the grid points.
Letting i,j be coordinates in a regular 2D grid, it is easy to see that the pivot on grid point (i, j)
is only determined by pivots on points (i − 1, j) and (i, j − 1). If there are n points on each of m
grid lines, we get the following generating relations for the pivots:
di,i =
a1,1 if i = 1
ai,i − ai,i−1d−1
i−1ai−1,i if 1 < i ≤ n
ai,i − ai,i−nd−1
i−nai−n,i if i = kn + 1 with k ≥ 1
ai,i − ai,i−1d−1
i−1ai−1,i
− ai,i−nd−1
i−nai−n,i
otherwise.
Conversely, we can describe the factorization algorithmically as
Initially: di,i = ai,i for all i
for i = 1..nm do:
di+1,i+1 = di+1,i+1 − ai+1,id−1
i,i ai,i+1 if there is no k
such that i = kn
di+n,i+n = di+n,i+n − ai+n,id−1
i,i ai,i+n if i + n ≤ nm
In the above we have assumed that the variables in the problem are ordered according to the
so-called “natural ordering”: a sequential numbering of the grid lines and the points within each
grid line. Below we will encounter different orderings of the variables.
Modified incomplete factorizations
One modification to the basic idea of incomplete factorizations is as follows: If the product
ai,ka−1
k,kak,j is nonzero, and fill is not allowed in position (i, j), instead of simply discarding this
fill quantity subtract it from the diagonal element ai,i. Such a factorization scheme is usually
called a “modified incomplete factorization”.
Mathematically this corresponds to forcing the preconditioner to have the same rowsums as
the original matrix. One reason for considering modified incomplete factorizations is the behavior
of the spectral condition number of the preconditioned system. It was mentioned above that for
second order elliptic equations the condition number of the coefficient matrix is O(h−2
) as a func-
tion of the discretization mesh width. This order of magnitude is preserved by simple incomplete
factorizations, although usually a reduction by a large constant factor is obtained.
Modified factorizations are of interest because, in combination with small perturbations, the
spectral condition number of the preconditioned system can be of a lower order. It was first proved
by Dupont, Kendall and Rachford [80] that a modified incomplete factorization of A + O(h2
)DA
gives κ(M−1
A) = O(h−1
) for the central difference case. More general proofs are given by
Gustafsson [111], Axelsson and Barker [14, §7.2], and Beauwens [31, 32].](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-51-320.jpg)
![42 CHAPTER 3. PRECONDITIONERS
for k = 1, . . . , 2n − 1
do in parallel for i = max(1, k + 1 − n), min(n, k)
j = k − i + 1
xi+(j−1)n ← Di+(j−1)nyi+(j−1)n
−Li+(j−1)ni − 1 + (j − 1)nxi−1+(j−1)n
−Li+(j−1)ni + (j − 2)nxi+(j−2)n
Figure 3.4: Wavefront solution of (D + L)x = u from a central difference problem on a domain
of n × n points.
Instead of keeping row sums constant, one can also keep column sums constant. In compu-
tational fluid mechanics this idea is justified with the argument that the material balance stays
constant over all iterates. (Equivalently, one wishes to avoid ‘artificial diffusion’.) Appleyard and
Cheshire [4] observed that if A and M have the same column sums, the choice
x0 = M−1
b
guarantees that the sum of the elements in r0 (the material balance error) is zero, and that all further
ri have elements summing to zero.
Modified incomplete factorizations can break down, especially when the variables are num-
bered other than in the natural row-by-row ordering. This was noted by Chan and Kuo [50], and a
full analysis was given by Eijkhout [85] and Notay [160].
A slight variant of modified incomplete factorizations consists of the class of “relaxed incom-
plete factorizations”. Here the fill is multiplied by a parameter 0 < α < 1 before it is subtracted
from the diagonal; see Ashcraft and Grimes [11], Axelsson and Lindskog [18, 19], Chan [44],
Eijkhout [85], Notay [161], Stone [193], and Van der Vorst [202]. For the dangers of MILU in the
presence of rounding error, see Van der Vorst [204].
Vectorization of the preconditioner solve
At first it may appear that the sequential time of solving a factorization is of the order of the number
of variables, but things are not quite that bad. Consider the special case of central differences on a
regular domain of n × n points. The variables on any diagonal in the domain, that is, in locations
(i, j) with i + j = k, depend only on those on the previous diagonal, that is, with i + j = k − 1.
Therefore it is possible to process the operations on such a diagonal, or ‘wavefront’, in parallel (see
figure 3.4), or have a vector computer pipeline them; see Van der Vorst [201, 203].
Another way of vectorizing the solution of the triangular factors is to use some form of expan-
sion of the inverses of the factors. Consider for a moment a lower triangular matrix, normalized
to the form I − L where L is strictly lower triangular). Its inverse can be given as either of the
following two series:
(I − L)−1
=
I + L + L2
+ L3
+ · · ·
(I + L)(I + L2
)(I + L4
) · · ·
(3.2)
(The first series is called a “Neumann expansion”, the second an “Euler expansion”. Both series are
finite, but their length prohibits practical use of this fact.) Parallel or vectorizable preconditioners
can be derived from an incomplete factorization by taking a small number of terms in either series.
Experiments indicate that a small number of terms, while giving high execution rates, yields almost
the full precision of the more recursive triangular solution (see Axelsson and Eijkhout [15] and
Van der Vorst [199]).](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-52-320.jpg)
![3.4. INCOMPLETE FACTORIZATION PRECONDITIONERS 43
Let M = (I + L)D(I + U) and y be given.
t ← y
for k = 1, . . . , p
t ← y − Lt
x ← D−1
t, t ← x
for k = 1, . . . , p
x ← t − Ux
Figure 3.5: Preconditioning step algorithm for a Neumann expansion M(p)
≈ M−1
of an incom-
plete factorization M = (I + L)D(I + U).
There are some practical considerations in implementing these expansion algorithms. For in-
stance, because of the normalization the L in equation (3.2) is not LA. Rather, if we have a precon-
ditioner (as described in section 3.4.2) described by
A = DA + LA + UA, M = (D + LA)D−1
(D + UA),
then we write
M = (I + L)D(I + U), where L = LAD−1
, U = D−1
UA.
Now we can choose whether or not to store the product LAD−1
. Doing so doubles the storage
requirements for the matrix, not doing so means that separate multiplications by LA and D−1
have
to be performed in the expansion.
Suppose then that the products L = LAD−1
and U = D−1
UA have been stored. We then
define M(p)
≈ M−1
by
M(p)
=
p
k=0
(−U)p
D−1
p
k=0
(−L)p
, (3.3)
and replace solving a system Mx = y for x by computing x = M(p)
y. This algorithm is given in
figure 3.5.
Parallelizing the preconditioner solve
The algorithms for vectorization outlined above can be used on parallel computers. For instance,
variables on a wavefront can be processed in parallel, by dividing the wavefront over processors.
More radical approaches for increasing the parallelism in incomplete factorizations are based on
a renumbering of the problem variables. For instance, on rectangular domains one could start
numbering the variables from all four corners simultaneously, thereby creating four simultaneous
wavefronts, and therefore four-fold parallelism (see Dongarra, et al. [70], Van der Vorst [200, 202]).
The most extreme case is the red/black ordering (or for more general matrices the multi-color
ordering) which gives the absolute minimum number of sequential steps.
Multi-coloring is also an attractive method for vector computers. Since points of one color
are uncoupled, they can be processed as one vector; see Doi [67], Melhem [153], and Poole and
Ortega [175].
However, for such ordering strategies there is usually a trade-off between the degree of paral-
lelism and the resulting number of iterations. The reason for this is that a different ordering may
give rise to a different error matrix, in particular the norm of the error matrix may vary considerably
between orderings. See experimental results by Duff and Meurant [78] and a partial explanation of
them by Eijkhout [84].](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-53-320.jpg)
![44 CHAPTER 3. PRECONDITIONERS
for i = 1, 2, . . .
Xi ← Aii
for i = 1, 2, . . .
Let Yi ≈ X−1
i
for j = i + 1, i + 2, . . .
if Aij = 0 and Aji = 0,
then Xj ← Xj − AjiYiAij
Figure 3.6: Block version of a D-ILU factorization
3.4.3 Block factorization methods
We can also consider block variants of preconditioners for accelerated methods. Block methods
are normally feasible if the problem domain is a Cartesian product grid; in that case a natural
division in lines (or planes in the 3-dimensional case), can be used for blocking, though incomplete
factorizations are not as effective in the 3-dimensional case; see for instance Kettler [133]. In such
a blocking scheme for Cartesian product grids, both the size and number of the blocks increases
with the overall problem size.
The idea behind block factorizations
The starting point for an incomplete block factorization is a partitioning of the matrix, as mentioned
in §3.2.1. Then an incomplete factorization is performed using the matrix blocks as basic entities
(see Axelsson [12] and Concus, Golub and Meurant [56] as basic references).
The most important difference with point methods arises in the inversion of the pivot blocks.
Whereas inverting a scalar is easily done, in the block case two problems arise. First, inverting the
pivot block is likely to be a costly operation. Second, initially the diagonal blocks of the matrix
are likely to be be sparse and we would like to maintain this type of structure throughout the
factorization. Hence the need for approximations of inverses arises.
In addition to this, often fill-in in off-diagonal blocks is discarded altogether. Figure 3.6
describes an incomplete block factorization that is analogous to the D-ILU factorization (sec-
tion 3.4.2) in that it only updates the diagonal blocks.
As in the case of incomplete point factorizations, the existence of incomplete block methods is
guaranteed if the coefficient matrix is an M-matrix. For a general proof, see Axelsson [13].
Approximate inverses
In block factorizations a pivot block is generally forced to be sparse, typically of banded form,
and that we need an approximation to its inverse that has a similar structure. Furthermore, this
approximation should be easily computable, so we rule out the option of calculating the full inverse
and taking a banded part of it.
The simplest approximation to A−1
is the diagonal matrix D of the reciprocals of the diagonal
of A: di,i = 1/ai,i.
Other possibilities were considered by Axelsson and Eijkhout [15], Axelsson and Polman [21],
Concus, Golub and Meurant [56], Eijkhout and Vassilevski [89], Kolotilina and Yeremin [140],
and Meurant [154]. One particular example is given in figure 3.7. It has the attractive theoretical
property that, if the original matrix is symmetric positive definite and a factorization with positive
diagonal D can be made, the approximation to the inverse is again symmetric positive definite.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-54-320.jpg)
![3.4. INCOMPLETE FACTORIZATION PRECONDITIONERS 45
Let X be a banded matrix,
factor X = (I + L)D−1
(I + U),
let Y = (I − U)D(I − L)
Figure 3.7: Algorithm for approximating the inverse of a banded matrix
X1 = A1,1
for i ≥ 1
let Yi ≈ X−1
i
Xi+1 = Ai+1,i+1 − Ai+1,iYiAi,i+1
Figure 3.8: Incomplete block factorization of a block tridiagonal matrix
Banded approximations to the inverse of banded matrices have a theoretical justification. In
the context of partial differential equations the diagonal blocks of the coefficient matrix are usually
strongly diagonally dominant. For such matrices, the elements of the inverse have a size that is ex-
ponentially decreasing in their distance from the main diagonal. See Demko, Moss and Smith [64]
for a general proof, and Eijkhout and Polman [88] for a more detailed analysis in the M-matrix
case.
The special case of block tridiagonality
In many applications, a block tridiagonal structure can be found in the coefficient matrix. Examples
are problems on a 2D regular grid if the blocks correspond to lines of grid points, and problems
on a regular 3D grid, if the blocks correspond to planes of grid points. Even if such a block
tridiagonal structure does not arise naturally, it can be imposed by renumbering the variables in a
Cuthill-McKee ordering [59].
Such a matrix has incomplete block factorizations of a particularly simple nature: since no fill
can occur outside the diagonal blocks (Ai,i), all properties follow from our treatment of the pivot
blocks. The generating recurrence for the pivot blocks also takes a simple form, see figure 3.8.
After the factorization we are left with sequences of Xi block forming the pivots, and of Yi blocks
approximating their inverses.
Two types of incomplete block factorizations
One reason that block methods are of interest is that they are potentially more suitable for vector
computers and parallel architectures. Consider the block factorization
A = (D + L)D−1
(D + U) = (D + L)(I + D−1
U)
where D is the block diagonal matrix of pivot blocks,6
and L, U are the block lower and upper
triangle of the factorization; they coincide with LA, UQ in the case of a block tridiagonal matrix.
We can turn this into an incomplete factorization by replacing the block diagonal matrix of
pivots D by the block diagonal matrix of incomplete factorization pivots X = diag(Xi), giving
M = (X + L)(I + X−1
U)
6Writing (I + LAD−1)(D + UA) is equally valid, but in practice harder to implement.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-55-320.jpg)
![46 CHAPTER 3. PRECONDITIONERS
For factorizations of this type (which covers all methods in Concus, Golub and Meurant [56] and
Kolotilina and Yeremin [140]) solving a linear system means solving smaller systems with the Xi
matrices.
Alternatively, we can replace D by a the inverse of the block diagonal matrix of the approxi-
mations to the inverses of the pivots, Y = diag(Yi), giving
M = (Y −1
+ L)(I + Y U).
For this second type (which was discussed by Meurant [154], Axelsson and Polman [21] and Ax-
elsson and Eijkhout [15]) solving a system with M entails multiplying by the Yi blocks. Therefore,
the second type has a much higher potential for vectorizability. Unfortunately, such a factorization
is theoretically more troublesome; see the above references or Eijkhout and Vassilevski [89].
3.4.4 Blocking over systems of partial differential equations
If the physical problem has several variables per grid point, that is, if there are several coupled
partial differential equations, it is possible to introduce blocking in a natural way.
Blocking of the equations (which gives a small number of very large blocks) was used by
Axelsson and Gustafsson [17] for the equations of linear elasticity, and blocking of the variables
per node (which gives many very small blocks) was used by Aarden and Karlsson [1] for the
semiconductor equations. A systematic comparison of the two approaches was made by Bank, et
al. [26].
3.4.5 Incomplete LQ factorizations
Saad [183] proposes to construct an incomplete LQ factorization of a general sparse matrix. The
idea is to orthogonalize the rows of the matrix by a Gram-Schmidt process (note that in sparse
matrices, most rows are typically orthogonal already, so that standard Gram-Schmidt may be not so
bad as in general). Saad suggest dropping strategies for the fill-in produced in the orthogonalization
process. It turns out that the resulting incomplete L factor can be viewed as the incomplete Choleski
factor of the matrix AAT
. Experiments show that using L in a CG process for the normal equations:
L−1
AAT
L−T
y = b is effective for some relevant problems.
3.5 Polynomial preconditioners
So far, we have described preconditioners in only one of two classes: those that approximate the
coefficient matrix, and where linear systems with the preconditioner as coefficient matrix are easier
to solve than the original system. Polynomial preconditioners can be considered as members of the
second class of preconditioners: direct approximations of the inverse of the coefficient matrix.
Suppose that the coefficient matrix A of the linear system is normalized to the form A = I −B,
and that the spectral radius of B is less than one. Using the Neumann series, we can write the
inverse of A as A−1
=
∞
k=0 Bk
, so an approximation may be derived by truncating this infinite
series. Since the iterative methods we are considering are already based on the idea of applying
polynomials in the coefficient matrix to the initial residual, there are analytic connections between
the basic method and polynomially accelerated one.
Dubois, Greenbaum and Rodrigue [76] investigated the relationship between a basic method
using a splitting A = M − N, and a polynomially preconditioned method with
M−1
p = (
p−1
i=0
(I − M−1
A)i
)M−1
.
The basic result is that for classical methods, k steps of the polynomially preconditioned method
are exactly equivalent to kp steps of the original method; for accelerated methods, specifically the](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-56-320.jpg)
![3.6. OTHER PRECONDITIONERS 47
Chebyshev method, the preconditioned iteration can improve the number of iterations by at most a
factor of p.
Although there is no gain in the number of times the coefficient matrix is applied, polynomial
preconditioning does eliminate a large fraction of the inner products and update operations, so there
may be an overall increase in efficiency.
Let us define a polynomial preconditioner more abstractly as any polynomial M = Pn(A)
normalized to P(0) = 1. Now the choice of the best polynomial preconditioner becomes that of
choosing the best polynomial that minimizes I − M−1
A . For the choice of the infinity norm we
thus obtain Chebyshev polynomials, and they require estimates of both a lower and upper bound
on the spectrum of A. These estimates may be derived from the conjugate gradient iteration itself;
see §5.1.
Since an accurate lower bound on the spectrum of A may be hard to obtain, Johnson, Micchelli
and Paul [125] and Saad [182] propose least squares polynomials based on several weight functions.
These functions only require an upper bound and this is easily computed, using for instance the
“Gerschgorin bound” maxi j |Ai,j|; see [209, §1.4]. Experiments comparing Chebyshev and
least squares polynomials can be found in Ashby, Manteuffel and Otto [8].
Application of polynomial preconditioning to symmetric indefinite problems is described by
Ashby, Manteuffel and Saylor [9]. There the polynomial is chosen so that it transforms the system
into a definite one.
3.6 Preconditioners from properties of the differential equation
A number of preconditioners exist that derive their justification from properties of the underlying
partial differential equation. We will cover some of them here (see also §5.5 and §5.4). These
preconditioners usually involve more work than the types discussed above, however, they allow for
specialized faster solution methods.
3.6.1 Preconditioning by the symmetric part
In §2.3.4 we pointed out that conjugate gradient methods for non-selfadjoint systems require the
storage of previously calculated vectors. Therefore it is somewhat remarkable that preconditioning
by the symmetric part (A + AT
)/2 of the coefficient matrix A leads to a method that does not need
this extended storage. Such a method was proposed by Concus and Golub [55] and Widlund [214].
However, solving a system with the symmetric part of a matrix may be no easier than solving
a system with the full matrix. This problem may be tackled by imposing a nested iterative method,
where a preconditioner based on the symmetric part is used. Vassilevski [210] proved that the
efficiency of this preconditioner for the symmetric part carries over to the outer method.
3.6.2 The use of fast solvers
In many applications, the coefficient matrix is symmetric and positive definite. The reason for this
is usually that the partial differential operator from which it is derived is self-adjoint, coercive,
and bounded (see Axelsson and Barker [14, §3.2]). It follows that for the coefficient matrix A the
following relation holds for any matrix B from a similar differential equation:
c1 ≤
xT
Ax
xT Bx
≤ c2 for all x,
where c1, c2 do not depend on the matrix size. The importance of this is that the use of B as a
preconditioner gives an iterative method with a number of iterations that does not depend on the
matrix size.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-57-320.jpg)
![48 CHAPTER 3. PRECONDITIONERS
Thus we can precondition our original matrix by one derived from a different PDE, if one can be
found that has attractive properties as preconditioner. The most common choice is to take a matrix
from a separable PDE. A system involving such a matrix can be solved with various so-called
“fast solvers”, such as FFT methods, cyclic reduction, or the generalized marching algorithm (see
Dorr [74], Swarztrauber [194], Bank [25] and Bank and Rose [27]).
As a simplest example, any elliptic operator can be preconditioned with the Poisson operator,
giving the iterative method
−∆(un+1 − un) = −(Lun − f).
In Concus and Golub [58] a transformation of this method is considered to speed up the conver-
gence. As another example, if the original matrix arises from
−(a(x, y)ux)x − (b(x, y)uy)y = f,
the preconditioner can be formed from
−(˜a(x)ux)x − (˜b(y)uy)y = f.
An extension to the non-self adjoint case is considered by Elman and Schultz [93].
Fast solvers are attractive in that the number of operations they require is (slightly higher than)
of the order of the number of variables. Coupled with the fact that the number of iterations in the
resulting preconditioned iterative methods is independent of the matrix size, such methods are close
to optimal. However, fast solvers are usually only applicable if the physical domain is a rectangle
or other Cartesian product structure. (For a domain consisting of a number of such pieces, domain
decomposition methods can be used; see §5.4).
3.6.3 Alternating Direction Implicit methods
The Poisson differential operator can be split in a natural way as the sum of two operators:
L = L1 + L2, where L1 = − ∂2
∂x2 , L2 = − ∂2
∂y2 .
Now let L1, L2 be discretized representations of L1, L2. Based on the observation that L1 + L2 =
(I + L1)(I + L2) − I − L1L2, iterative schemes such as
(1 + αL1)(1 + αL2)u(m+1)
= [(1 + βL1)(1 + βL2)] u(m)
with suitable choices of α and β have been proposed.
This alternating direction implicit, or ADI, method was first proposed as a solution method
for parabolic equations. The u(m)
are then approximations on subsequent time steps. However,
it can also be used for the steady state, that is, for solving elliptic equations. In that case, the
u(m)
become subsequent iterates; see D’Yakonov [81], Fairweather, Gourlay and Mitchell [96],
Hadjidimos [118], and Peaceman and Rachford [172]. Generalization of this scheme to variable
coefficients or fourth order elliptic problems is relatively straightforward.
The above method is implicit since it requires systems solutions, and it alternates the x and y
(and if necessary z) directions. It is attractive from a practical point of view (although mostly on
tensor product grids), since solving a system with, for instance, a matrix I + αL1 entails only a
number of uncoupled tridiagonal solutions. These need very little storage over that needed for the
matrix, and they can be executed in parallel, or one can vectorize over them.
A theoretical reason that ADI preconditioners are of interest is that they can be shown to be
spectrally equivalent to the original coefficient matrix. Hence the number of iterations is bounded
independent of the condition number.
However, there is a problem of data distribution. For vector computers, either the system solu-
tion with L1 or with L2 will involve very large strides: if columns of variables in the grid are stored](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-58-320.jpg)
![Chapter 4
Related Issues
4.1 Complex Systems
Conjugate gradient methods for real symmetric systems can be applied to complex Hermitian sys-
tems in a straightforward manner. For non-Hermitian complex systems we distinguish two cases.
In general, for any coefficient matrix a CGNE method is possible, that is, a conjugate gradients
method on the normal equations AH
Ax = AH
b, or one can split the system into real and complex
parts and use a method such as GMRES on the resulting real nonsymmetric system. However, in
certain practical situations the complex system is non-Hermitian but symmetric.
Complex symmetric systems can be solved by a classical conjugate gradient or Lanczos
method, that is, with short recurrences, if the complex inner product (x, y) = ¯xT
y is replaced by
(x, y) = xT
y. Like the BiConjugate Gradient method, this method is susceptible to breakdown,
that is, it can happen that xT
x = 0 for x = 0. A look-ahead strategy can remedy this in most
cases (see Freund [99] and Van der Vorst and Melissen [206]).
4.2 Stopping Criteria
An iterative method produces a sequence {x(i)
} of vectors converging to the vector x satisfying
the n × n system Ax = b. To be effective, a method must decide when to stop. A good stopping
criterion should
1. identify when the error e(i)
≡ x(i)
− x is small enough to stop,
2. stop if the error is no longer decreasing or decreasing too slowly, and
3. limit the maximum amount of time spent iterating.
For the user wishing to read as little as possible, the following simple stopping criterion will
likely be adequate. The user must supply the quantities maxit, b , stop tol, and preferably also
A :
• The integer maxit is the maximum number of iterations the algorithm will be permitted to
perform.
• The real number A is a norm of A. Any reasonable (order of magnitude) approximation
of the absolute value of the largest entry of the matrix A will do.
• The real number b is a norm of b. Again, any reasonable approximation of the absolute
value of the largest entry of b will do.
51](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-61-320.jpg)
![4.2. STOPPING CRITERIA 53
consistent pair of the above. ( x 2 and A F , as well as x 2 and A 2, both form mutually
consistent pairs.) All these norms satisfy the triangle inequality x + y ≤ x + y and A +
B ≤ A + B , as well as Ax ≤ A · x for mutually consistent pairs. (For more details
on the properties of norms, see Golub and Van Loan [108].)
One difference between these norms is their dependence on dimension. A vector x of length n
with entries uniformly distributed between 0 and 1 will satisfy x ∞ ≤ 1, but x 2 will grow like√
n and x 1 will grow like n. Therefore a stopping criterion based on x 1 (or x 2) may have
to be permitted to grow proportional to n (or
√
n) in order that it does not become much harder to
satisfy for large n.
There are two approaches to bounding the inaccuracy of the computed solution to Ax = b.
Since e(i)
, which we will call the forward error, is hard to estimate directly, we introduce the
backward error, which allows us to bound the forward error. The normwise backward error is de-
fined as the smallest possible value of max{ δA / A , δb / b } where x(i)
is the exact solution
of (A + δA)x(i)
= (b + δb) (here δA denotes a general matrix, not δ times A; the same goes for
δb). The backward error may be easily computed from the residual r(i)
= Ax(i)
−b; we show how
below. Provided one has some bound on the inverse of A, one can bound the forward error in terms
of the backward error via the simple equality
e(i)
= x(i)
− x = A−1
(Ax(i)
− b) = A−1
r(i)
,
which implies e(i)
≤ A−1
· r(i)
. Therefore, a stopping criterion of the form “stop when
r(i)
≤ τ” also yields an upper bound on the forward error e(i)
≤ τ · A−1
. (Sometimes we
may prefer to use the stricter but harder to estimate bound e(i)
≤ |A−1
| · |r(i)
| ; see §4.2.3.
Here |X| is the matrix or vector of absolute values of components of X.)
The backward error also has a direct interpretation as a stopping criterion, in addition to
supplying a bound on the forward error. Recall that the backward error is the smallest change
max{ δA / A , δb / b } to the problem Ax = b that makes x(i)
an exact solution of (A +
δA)x(i)
= b + δb. If the original data A and b have errors from previous computations or mea-
surements, then it is usually not worth iterating until δA and δb are even smaller than these errors.
For example, if the machine precision is ε, it is not worth making δA ≤ ε A and δb ≤ ε b ,
because just rounding the entries of A and b to fit in the machine creates errors this large.
Based on this discussion, we will now consider some stopping criteria and their properties.
Above we already mentioned
Criterion 1. r(i)
≤ S1 ≡ stop tol · ( A · x(i)
+ b ). This is equivalent to asking that
the backward error δA and δb described above satisfy δA ≤ stop tol · A and δb ≤
stop tol · b . This criterion yields the forward error bound
e(i)
≤ A−1
· r(i)
≤ stop tol · A−1
· ( A · x(i)
+ b ) .
The second stopping criterion we discussed, which does not require A , may be much more
stringent than Criterion 1:
Criterion 2. r(i)
≤ S2 ≡ stop tol · b . This is equivalent to asking that the backward error
δA and δb satisfy δA = 0 and δb ≤ tol · b . One difficulty with this method is that if
A · x b , which can only occur if A is very ill-conditioned and x nearly lies in the
null space of A, then it may be difficult for any method to satisfy the stopping criterion. To
see that A must be very ill-conditioned, note that
1
A · x
b
=
A · A−1
b
b
≤ A · A−1
.
This criterion yields the forward error bound
e(i)
≤ A−1
· r(i)
≤ stop tol · A−1
· b](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-63-320.jpg)
![4.2. STOPPING CRITERIA 55
4.2.2 When r(i)
or r(i)
is not readily available
It is possible to design an iterative algorithm for which r(i)
= Ax(i)
− b or r(i)
is not directly
available, although this is not the case for any algorithms in this book. For completeness, however,
we discuss stopping criteria in this case.
For example, if ones “splits” A = M − N to get the iterative method x(i)
= M−1
Nx(i−1)
+
M−1
b ≡ Gx(i−1)
+ c, then the natural residual to compute is ˆr(i)
= x(i)
− Gx(i)
− c =
M−1
(Ax(i)
− b) = M−1
r(i)
. In other words, the residual ˆr(i)
is the same as the residual
of the preconditioned system M−1
Ax = M−1
b. In this case, it is hard to interpret ˆr(i)
as
a backward error for the original system Ax = b, so we may instead derive a forward error
bound e(i)
= A−1
M ˆr(i)
≤ A−1
M · ˆr(i)
. Using this as a stopping criterion requires
an estimate of A−1
M . In the case of methods based on splitting A = M − N, we have
A−1
M = (M − N)−1
M = (I − G)−1
, and A−1
M = (I − G)−1
≤ 1/(1 − G ).
Another example is an implementation of the preconditioned conjugate gradient algorithm
which computes r(i)
M−1/2,2 = (r(i)T
M−1
r(i)
)1/2
instead of r(i)
2 (the implementation
in this book computes the latter). Such an implementation could use the stopping criterion
r(i)
M−1/2,2/ r(0)
M−1/2,2 ≤ tol as in Criterion 5. We may also use it to get the forward error
bound e(i)
≤ A−1
M1/2
· r(i)
M−1/2,2, which could also be used in a stopping criterion.
4.2.3 Estimating A−1
Bounds on the error e(i)
inevitably rely on bounds for A−1
, since e(i)
= A−1
r(i)
. There is a
large number of problem dependent ways to estimate A−1
; we mention a few here.
When a splitting A = M − N is used to get an iteration
x(i)
= M−1
Nx(i−1)
+ M−1
b = Gx(i−1)
+ c,
then the matrix whose inverse norm we need is I − G. Often, we know how to estimate G if the
splitting is a standard one such as Jacobi or SOR, and the matrix A has special characteristics such
as Property A. Then we may estimate (I − G)−1
≤ 1/(1 − G ).
When A is symmetric positive definite, and Chebyshev acceleration with adaptation of param-
eters is being used, then at each step the algorithm estimates the largest and smallest eigenvalues
λmax(A) and λmin(A) of A anyway. Since A is symmetric positive definite, A−1
2 = λ−1
min(A).
This adaptive estimation is often done using the Lanczos algorithm (see section 5.1), which can
usually provide good estimates of the largest (rightmost) and smallest (leftmost) eigenvalues of a
symmetric matrix at the cost of a few matrix-vector multiplies. For general nonsymmetric A, we
may apply the Lanczos method to AAT
or AT
A, and use the fact that A−1
2 = 1/λ
1/2
min(AAT
) =
1/λ
1/2
min(AT
A).
It is also possible to estimate A−1
∞ provided one is willing to solve a few systems of linear
equations with A and AT
as coefficient matrices. This is often done with dense linear system
solvers, because the extra cost of these systems is O(n2
), which is small compared to the cost
O(n3
) of the LU decomposition (see Hager [120], Higham [123] and Anderson, et al. [3]). This is
not the case for iterative solvers, where the cost of these solves may well be several times as much
as the original linear system. Still, if many linear systems with the same coefficient matrix and
differing right-hand-sides are to be solved, it is a viable method.
The approach in the last paragraph also lets us estimate the alternate error bound e(i)
∞ ≤
|A−1
| · |r(i)
| ∞. This may be much smaller than the simpler A−1
∞ · r(i)
∞ in the case
where the rows of A are badly scaled; consider the case of a diagonal matrix A with widely varying
diagonal entries. To compute |A−1
| · |r(i)
| ∞, let R denote the diagonal matrix with diagonal
entries equal to the entries of |r(i)
|; then |A−1
| · |r(i)
| ∞ = A−1
R ∞ (see Arioli, Demmel and
Duff [5]). A−1
R ∞ can be estimated using the technique in the last paragraph since multiplying
by A−1
R or (A−1
R)T
= RT
A−T
is no harder than multiplying by A−1
and A−T
and also by R,
a diagonal matrix.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-65-320.jpg)
![56 CHAPTER 4. RELATED ISSUES
4.2.4 Stopping when progress is no longer being made
In addition to limiting the total amount of work by limiting the maximum number of iterations one
is willing to do, it is also natural to consider stopping when no apparent progress is being made.
Some methods, such as Jacobi and SOR, often exhibit nearly monotone linear convergence, at least
after some initial transients, so it is easy to recognize when convergence degrades. Other methods,
like the conjugate gradient method, exhibit “plateaus” in their convergence, with the residual norm
stagnating at a constant value for many iterations before decreasing again; in principle there can
be many such plateaus (see Greenbaum and Strakos [109]) depending on the problem. Still other
methods, such as CGS, can appear wildly nonconvergent for a large number of steps before the
residual begins to decrease; convergence may continue to be erratic from step to step.
In other words, while it is a good idea to have a criterion that stops when progress towards a
solution is no longer being made, the form of such a criterion is both method and problem depen-
dent.
4.2.5 Accounting for floating point errors
The error bounds discussed in this section are subject to floating point errors, most of which are
innocuous, but which deserve some discussion.
The infinity norm x ∞ = maxj |xj| requires the fewest floating point operations to compute,
and cannot overflow or cause other exceptions if the xj are themselves finite2
. On the other hand,
computing x 2 = ( j |xj|2
)1/2
in the most straightforward manner can easily overflow or lose
accuracy to underflow even when the true result is far from either the overflow or underflow thresh-
olds. For this reason, a careful implementation for computing x 2 without this danger is available
(subroutine snrm2 in the BLAS [71] [143]), but it is more expensive than computing x ∞.
Now consider computing the residual r(i)
= Ax(i)
− b by forming the matrix-vector product
Ax(i)
and then subtracting b, all in floating point arithmetic with relative precision ε. A standard
error analysis shows that the error δr(i)
in the computed r(i)
is bounded by δr(i)
≤ O(ε)( A ·
x(i)
+ b ), where O(ε) is typically bounded by nε, and usually closer to
√
nε. This is why
one should not choose stop tol ≤ ε in Criterion 1, and why Criterion 2 may not be satisfied by any
method. This uncertainty in the value of r(i)
induces an uncertainty in the error e(i)
= A−1
r(i)
of
at most O(ε) A−1
· ( A · x(i)
+ b ). A more refined bound is that the error (δr(i)
)j in the
jth component of r(i)
is bounded by O(ε) times the jth component of |A| · |x(i)
| + |b|, or more
tersely |δr(i)
| ≤ O(ε)(|A| · |x(i)
| + |b|). This means the uncertainty in e(i)
is really bounded by
O(ε) |A−1
|·(|A|·|x(i)
|+|b|) . This last quantity can be estimated inexpensively provided solving
systems with A and AT
as coefficient matrices is inexpensive (see the last paragraph of §4.2.3).
Both these bounds can be severe overestimates of the uncertainty in e(i)
, but examples exist where
they are attainable.
4.3 Data Structures
The efficiency of any of the iterative methods considered in previous sections is determined primar-
ily by the performance of the matrix-vector product and the preconditioner solve, and therefore on
the storage scheme used for the matrix and the preconditioner. Since iterative methods are typically
used on sparse matrices, we will review here a number of sparse storage formats. Often, the storage
scheme used arises naturally from the specific application problem.
In this section we will review some of the more popular sparse matrix formats that are used in
numerical software packages such as ITPACK [139] and NSPCG [164]. After surveying the various
formats, we demonstrate how the matrix-vector product and an incomplete factorization solve are
formulated using two of the sparse matrix formats.
2IEEE standard floating point arithmetic permits computations with ±∞ and NaN, or Not-a-Number, symbols.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-66-320.jpg)
![4.3. DATA STRUCTURES 57
4.3.1 Survey of Sparse Matrix Storage Formats
If the coefficient matrix A is sparse, large-scale linear systems of the form Ax = b can be most
efficiently solved if the zero elements of A are not stored. Sparse storage schemes allocate con-
tiguous storage in memory for the nonzero elements of the matrix, and perhaps a limited number
of zeros. This, of course, requires a scheme for knowing where the elements fit into the full matrix.
There are many methods for storing the data (see for instance Saad [185] and Eijkhout [86]).
Here we will discuss Compressed Row and Column Storage, Block Compressed Row Storage,
Diagonal Storage, Jagged Diagonal Storage, and Skyline Storage.
Compressed Row Storage (CRS)
The Compressed Row and Column (in the next section) Storage formats are the most general: they
make absolutely no assumptions about the sparsity structure of the matrix, and they don’t store any
unnecessary elements. On the other hand, they are not very efficient, needing an indirect addressing
step for every single scalar operation in a matrix-vector product or preconditioner solve.
The Compressed Row Storage (CRS) format puts the subsequent nonzeros of the matrix rows
in contiguous memory locations. Assuming we have a nonsymmetric sparse matrix A, we cre-
ate 3 vectors: one for floating-point numbers (val), and the other two for integers (col ind,
row ptr). The val vector stores the values of the nonzero elements of the matrix A, as they
are traversed in a row-wise fashion. The col ind vector stores the column indexes of the el-
ements in the val vector. That is, if val(k) = ai,j then col ind(k) = j. The row ptr
vector stores the locations in the val vector that start a row, that is, if val(k) = ai,j then
row ptr(i) ≤ k < row ptr(i + 1). By convention, we define row ptr(n + 1) = nnz + 1,
where nnz is the number of nonzeros in the matrix A. The storage savings for this approach is
significant. Instead of storing n2
elements, we need only 2nnz + n + 1 storage locations.
As an example, consider the nonsymmetric matrix A defined by
A =
10 0 0 0 −2 0
3 9 0 0 0 3
0 7 8 7 0 0
3 0 8 7 5 0
0 8 0 9 9 13
0 4 0 0 2 −1
. (4.1)
The CRS format for this matrix is then specified by the arrays {val, col ind, row ptr}
given below
val 10 -2 3 9 3 7 8 7 3 · · · 9 13 4 2 -1
col ind 1 5 1 2 6 2 3 4 1 · · · 5 6 2 5 6
row ptr 1 3 6 9 13 17 20 .
If the matrix A is symmetric, we need only store the upper (or lower) triangular portion of the
matrix. The trade-off is a more complicated algorithm with a somewhat different pattern of data
access.
Compressed Column Storage (CCS)
Analogous to Compressed Row Storage there is Compressed Column Storage (CCS), which is also
called the Harwell-Boeing sparse matrix format [77]. The CCS format is identical to the CRS
format except that the columns of A are stored (traversed) instead of the rows. In other words, the
CCS format is the CRS format for AT
.
The CCS format is specified by the 3 arrays {val, row ind, col ptr}, where row ind
stores the row indices of each nonzero, and col ptr stores the index of the elements in val
which start a column of A. The CCS format for the matrix A in (4.1) is given by](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-67-320.jpg)
![58 CHAPTER 4. RELATED ISSUES
val 10 3 3 9 7 8 4 8 8 · · · 9 2 3 13 -1
row ind 1 2 4 2 3 5 6 3 4 · · · 5 6 2 5 6
col ptr 1 4 8 10 13 17 20 .
Block Compressed Row Storage (BCRS)
If the sparse matrix A is comprised of square dense blocks of nonzeros in some regular pattern,
we can modify the CRS (or CCS) format to exploit such block patterns. Block matrices typically
arise from the discretization of partial differential equations in which there are several degrees of
freedom associated with a grid point. We then partition the matrix in small blocks with a size equal
to the number of degrees of freedom, and treat each block as a dense matrix, even though it may
have some zeros.
If nb is the dimension of each block and nnzb is the number of nonzero blocks in the n × n
matrix A, then the total storage needed is nnz = nnzb × n2
b. The block dimension nd of A is then
defined by nd = n/nb.
Similar to the CRS format, we require 3 arrays for the BCRS format: a rectangular array for
floating-point numbers ( val(1 : nnzb,1 : nb,1 : nb)) which stores the nonzero blocks in
(block) row-wise fashion, an integer array (col ind(1 : nnzb)) which stores the actual column
indices in the original matrix A of the (1, 1) elements of the nonzero blocks, and a pointer array
(row blk(1 : nd + 1)) whose entries point to the beginning of each block row in val(:,:,:)
and col ind(:). The savings in storage locations and reduction in indirect addressing for BCRS
over CRS can be significant for matrices with a large nb.
Compressed Diagonal Storage (CDS)
If the matrix A is banded with bandwidth that is fairly constant from row to row, then it is worth-
while to take advantage of this structure in the storage scheme by storing subdiagonals of the
matrix in consecutive locations. Not only can we eliminate the vector identifying the column and
row, we can pack the nonzero elements in such a way as to make the matrix-vector product more
efficient. This storage scheme is particularly useful if the matrix arises from a finite element or
finite difference discretization on a tensor product grid.
We say that the matrix A = (ai,j) is banded if there are nonnegative constants p, q, called the
left and right halfbandwidth, such that ai,j = 0 only if i − p ≤ j ≤ i + q. In this case, we can
allocate for the matrix A an array val(1:n,-p:q). The declaration with reversed dimensions
(-p:q,n) corresponds to the LINPACK band format [72], which unlike CDS, does not allow for
an efficiently vectorizable matrix-vector multiplication if p + q is small.
Usually, band formats involve storing some zeros. The CDS format may even contain some ar-
ray elements that do not correspond to matrix elements at all. Consider the nonsymmetric matrix A
defined by
A =
10 −3 0 0 0 0
3 9 6 0 0 0
0 7 8 7 0 0
0 0 8 7 5 0
0 0 0 9 9 13
0 0 0 0 2 −1
. (4.2)
Using the CDS format, we store this matrix A in an array of dimension (6,-1:1) using the
mapping
val(i, j) = ai,i+j. (4.3)
Hence, the rows of the val(:,:) array are](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-68-320.jpg)
![4.3. DATA STRUCTURES 59
val(:,-1) 0 3 7 8 9 2
val(:, 0) 10 9 8 7 9 -1
val(:,+1) -3 6 7 5 13 0
.
Notice the two zeros corresponding to non-existing matrix elements.
A generalization of the CDS format more suitable for manipulating general sparse matrices on
vector supercomputers is discussed by Melhem in [153]. This variant of CDS uses a stripe data
structure to store the matrix A. This structure is more efficient in storage in the case of varying
bandwidth, but it makes the matrix-vector product slightly more expensive, as it involves a gather
operation.
As defined in [153], a stripe in the n × n matrix A is a set of positions S = {(i, σ(i)); i ∈ I ⊆
In}, where In = {1, . . . , n} and σ is a strictly increasing function. Specifically, if (i, σ(i)) and
(j, σ(j)) are in S, then
i < j → σ(i) < σ(j).
When computing the matrix-vector product y = Ax using stripes, each (i, σk(i)) element of A in
stripe Sk is multiplied with both xi and xσk(i) and these products are accumulated in yσk(i) and yi,
respectively. For the nonsymmetric matrix A defined by
A =
10 −3 0 1 0 0
0 9 6 0 −2 0
3 0 8 7 0 0
0 6 0 7 5 4
0 0 0 0 9 13
0 0 0 0 5 −1
, (4.4)
the 4 stripes of the matrix A stored in the rows of the val(:,:) array would be
val(:,-1) 0 0 3 6 0 5
val(:, 0) 10 9 8 7 9 -1
val(:,+1) 0 -3 6 7 5 13
val(:,+2) 0 1 -2 0 4 0
.
Jagged Diagonal Storage (JDS)
The Jagged Diagonal Storage format can be useful for the implementation of iterative methods on
parallel and vector processors (see Saad [184]). Like the Compressed Diagonal format, it gives a
vector length essentially of the size of the matrix. It is more space-efficient than CDS at the cost of
a gather/scatter operation.
A simplified form of JDS, called ITPACK storage or Purdue storage, can be described as fol-
lows. In the matrix from (4.4) all elements are shifted left:
10 −3 0 1 0 0
0 9 6 0 −2 0
3 0 8 7 0 0
0 6 0 7 5 4
0 0 0 0 9 13
0 0 0 0 5 −1
−→
10 −3 1
9 6 −2
3 8 7
6 7 5 4
9 13
5 −1
after which the columns are stored consecutively. All rows are padded with zeros on the right to
give them equal length. Corresponding to the array of matrix elements val(:,:), an array of
column indices, col ind(:,:) is also stored:
val(:, 1) 10 9 3 6 9 5
val(:, 2) −3 6 8 7 13 −1
val(:, 3) 1 −2 7 5 0 0
val(:, 4) 0 0 0 4 0 0
,](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-69-320.jpg)
![60 CHAPTER 4. RELATED ISSUES
col ind(:, 1) 1 2 1 2 5 5
col ind(:, 2) 2 3 3 4 6 6
col ind(:, 3) 4 5 4 5 0 0
col ind(:, 4) 0 0 0 6 0 0
.
It is clear that the padding zeros in this structure may be a disadvantage, especially if the
bandwidth of the matrix varies strongly. Therefore, in the CRS format, we reorder the rows of the
matrix decreasingly according to the number of nonzeros per row. The compressed and permuted
diagonals are then stored in a linear array. The new data structure is called jagged diagonals.
The number of jagged diagonals is equal to the number of nonzeros in the first row, i.e., the
largest number of nonzeros in any row of A. The data structure to represent the n × n matrix A
therefore consists of a permutation array (perm(1:n)) which reorders the rows, a floating-point
array (jdiag(:)) containing the jagged diagonals in succession, an integer array (col ind(:))
containing the corresponding column indices indices, and finally a pointer array (jd ptr(:))
whose elements point to the beginning of each jagged diagonal. The advantages of JDS for matrix
multiplications are discussed by Saad in [184].
The JDS format for the above matrix A in using the linear arrays {perm, jdiag, col ind,
jd ptr} is given below (jagged diagonals are separated by semicolons)
jdiag 1 3 7 8 10 2; 9 9 8 · · · -1; 9 6 7 5; 13
col ind 1 1 2 3 1 5; 4 2 3 · · · 6; 5 3 4 5; 6
perm 5 2 3 4 1 6 jd ptr 1 7 13 17 .
Skyline Storage (SKS)
The final storage scheme we consider is for skyline matrices, which are also called variable band
or profile matrices (see Duff, Erisman and Reid [79]). It is mostly of importance in direct solution
methods, but it can be used for handling the diagonal blocks in block matrix factorization methods.
A major advantage of solving linear systems having skyline coefficient matrices is that when piv-
oting is not necessary, the skyline structure is preserved during Gaussian elimination. If the matrix
is symmetric, we only store its lower triangular part. A straightforward approach in storing the el-
ements of a skyline matrix is to place all the rows (in order) into a floating-point array (val(:)),
and then keep an integer array (row ptr(:)) whose elements point to the beginning of each row.
The column indices of the nonzeros stored in val(:) are easily derived and are not stored.
For a nonsymmetric skyline matrix such as the one illustrated in Figure 4.1, we store the lower
triangular elements in SKS format, and store the upper triangular elements in a column-oriented
SKS format (transpose stored in row-wise SKS format). These two separated substructures can be
linked in a variety of ways. One approach, discussed by Saad in [185], is to store each row of the
lower triangular part and each column of the upper triangular part contiguously into the floating-
point array (val(:)). An additional pointer is then needed to determine where the diagonal
elements, which separate the lower triangular elements from the upper triangular elements, are
located.
4.3.2 Matrix vector products
In many of the iterative methods discussed earlier, both the product of a matrix and that of its
transpose times a vector are needed, that is, given an input vector x we want to compute products
y = Ax and y = AT
x.
We will present these algorithms for two of the storage formats from §4.3: CRS and CDS.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-70-320.jpg)
![4.3. DATA STRUCTURES 65
for i = 1, n
x_tmp(i) = x(i)
end;
for i = 1, n
z(i) = x_tmp(i)
tmp = pivots(i) * z(i)
for j = diag_ptr(i)+1, row_ptr(i+1)-1
x_tmp(col_ind(j)) = x_tmp(col_ind(j)) - tmp * val(j)
end;
end;
for i = n, 1 (step -1)
y(i) = pivots(i) * z(i)
for j = row_ptr(i), diag_ptr(i)-1
z(col_ind(j)) = z(col_ind(j)) - val(j) * y(i)
end;
end;
The extra temporary x tmp is used only for clarity, and can be overlapped with z. Both x tmp and
z can be considered to be equivalent to y. Overall, a CRS-based preconditioner solve uses short
vector lengths, indirect addressing, and has essentially the same memory traffic patterns as that of
the matrix-vector product.
Generating a CRS-based ILU(k) Incomplete Factorization
Incomplete factorizations with several levels of fill allowed are more accurate than the D-ILU
factorization described above. On the other hand, they require more storage, and are considerably
harder to implement (much of this section is based on algorithms for a full factorization of a sparse
matrix as found in Duff, Erisman and Reid [79]).
As a preliminary, we need an algorithm for adding two vectors x and y, both stored in sparse
storage. Let lx be the number of nonzero components in x, let x be stored in x, and let xind be
an integer array such that
if xind(j)=i then x(j) = xi.
Similarly, y is stored as ly, y, yind.
We now add x ← x + y by first copying y into a full vector w then adding w to x. The total
number of operations will be O(lx + ly)3
:
% copy y into w
for i=1,ly
w( yind(i) ) = y(i)
% add w to x wherever x is already nonzero
for i=1,lx
if w( xind(i) ) <> 0
x(i) = x(i) + w( xind(i) )
w( xind(i) ) = 0
% add w to x by creating new components
% wherever x is still zero
for i=1,ly
if w( yind(i) ) <> 0 then
lx = lx+1
3This is not counting the initial zeroing of the w array.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-75-320.jpg)
![4.4. PARALLELISM 69
1. The communication required for the reduction of the inner product for γ can be overlapped
with the update for x(i)
, (which could in fact have been done in the previous iteration step).
2. The reduction of the inner product for ρi+1 can be overlapped with the remaining part of the
preconditioning operation at the beginning of the next iteration.
3. The computation of a segment of p(i)
can be followed immediately by the computation of a
segment of q(i)
, and this can be followed by the computation of a part of the inner product.
This saves on load operations for segments of p(i)
and q(i)
.
For a more detailed discussion see Demmel, Heath and Van der Vorst [66]. This algorithm can
be extended trivially to preconditioners of LDLT
form, and nonsymmetric preconditioners in the
Biconjugate Gradient Method.
Fewer synchronization points
Several authors have found ways to eliminate some of the synchronization points induced by the
inner products in methods such as CG. One strategy has been to replace one of the two inner prod-
ucts typically present in conjugate gradient-like methods by one or two others in such a way that
all inner products can be performed simultaneously. The global communication can then be pack-
aged. A first such method was proposed by Saad [181] with a modification to improve its stability
suggested by Meurant [155]. Recently, related methods have been proposed by Chronopoulos and
Gear [54], D’Azevedo and Romine [61], and Eijkhout [87]. These schemes can also be applied to
nonsymmetric methods such as BiCG. The stability of such methods is discussed by D’Azevedo,
Eijkhout and Romine [60].
Another approach is to generate a number of successive Krylov vectors (see §2.3.4) and orthog-
onalize these as a block (see Van Rosendale [208], and Chronopoulos and Gear [54]).
4.4.2 Vector updates
Vector updates are trivially parallelizable: each processor updates its own segment.
4.4.3 Matrix-vector products
The matrix–vector products are often easily parallelized on shared-memory machines by split-
ting the matrix in strips corresponding to the vector segments. Each processor then computes the
matrix–vector product of one strip. For distributed-memory machines, there may be a problem if
each processor has only a segment of the vector in its memory. Depending on the bandwidth of
the matrix, we may need communication for other elements of the vector, which may lead to com-
munication bottlenecks. However, many sparse matrix problems arise from a network in which
only nearby nodes are connected. For example, matrices stemming from finite difference or finite
element problems typically involve only local connections: matrix element ai,j is nonzero only if
variables i and j are physically close. In such a case, it seems natural to subdivide the network, or
grid, into suitable blocks and to distribute them over the processors. When computing Api, each
processor requires the values of pi at some nodes in neighboring blocks. If the number of con-
nections to these neighboring blocks is small compared to the number of internal nodes, then the
communication time can be overlapped with computational work. For more detailed discussions on
implementation aspects for distributed memory systems, see De Sturler [62] and Pommerell [174].
4.4.4 Preconditioning
Preconditioning is often the most problematic part of parallelizing an iterative method. We will
mention a number of approaches to obtaining parallelism in preconditioning.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-79-320.jpg)
![70 CHAPTER 4. RELATED ISSUES
Discovering parallelism in sequential preconditioners. Certain preconditioners were not de-
veloped with parallelism in mind, but they can be executed in parallel. Some examples are domain
decomposition methods (see §5.4), which provide a high degree of coarse grained parallelism, and
polynomial preconditioners (see §3.5), which have the same parallelism as the matrix-vector prod-
uct.
Incomplete factorization preconditioners are usually much harder to parallelize: using wave-
fronts of independent computations (see for instance Paolini and Radicati di Brozolo [169]) a
modest amount of parallelism can be attained, but the implementation is complicated. For in-
stance, a central difference discretization on regular grids gives wavefronts that are hyperplanes
(see Van der Vorst [201, 203]).
More parallel variants of sequential preconditioners. Variants of existing sequential incom-
plete factorization preconditioners with a higher degree of parallelism have been devised, though
they are perhaps less efficient in purely scalar terms than their ancestors. Some examples are: re-
orderings of the variables (see Duff and Meurant [78] and Eijkhout [84]), expansion of the factors
in a truncated Neumann series (see Van der Vorst [199]), various block factorization methods (see
Axelsson and Eijkhout [15] and Axelsson and Polman [21]), and multicolor preconditioners.
Multicolor preconditioners have optimal parallelism among incomplete factorization methods,
since the minimal number of sequential steps equals the color number of the matrix graphs. For
theory and appplications to parallelism see Jones and Plassman [126, 127].
Fully decoupled preconditioners. If all processors execute their part of the preconditioner solve
without further communication, the overall method is technically a block Jacobi preconditioner
(see §3.2.1). While their parallel execution is very efficient, they may not be as effective as more
complicated, less parallel preconditioners, since improvement in the number of iterations may be
only modest. To get a bigger improvement while retaining the efficient parallel execution, Radi-
cati di Brozolo and Robert [177] suggest that one construct incomplete decompositions on slightly
overlapping domains. This requires communication similar to that for matrix–vector products.
4.4.5 Wavefronts in the Gauss-Seidel and Conjugate Gradient methods
At first sight, the Gauss-Seidel method (and the SOR method which has the same basic structure)
seems to be a fully sequential method. A more careful analysis, however, reveals a high degree of
parallelism if the method is applied to sparse matrices such as those arising from discretized partial
differential equations.
We start by partitioning the unknowns in wavefronts. The first wavefront contains those un-
knowns that (in the directed graph of D − L) have no predecessor; subsequent wavefronts are
then sets (this definition is not necessarily unique) of successors of elements of the previous wave-
front(s), such that no successor/predecessor relations hold among the elements of this set. It is clear
that all elements of a wavefront can be processed simultaneously, so the sequential time of solving
a system with D − L can be reduced to the number of wavefronts.
Next, we observe that the unknowns in a wavefront can be computed as soon as all wavefronts
containing its predecessors have been computed. Thus we can, in the absence of tests for con-
vergence, have components from several iterations being computed simultaneously. Adams and
Jordan [2] observe that in this way the natural ordering of unknowns gives an iterative method that
is mathematically equivalent to a multi-color ordering.
In the multi-color ordering, all wavefronts of the same color are processed simultaneously. This
reduces the number of sequential steps for solving the Gauss-Seidel matrix to the number of colors,
which is the smallest number d such that wavefront i contains no elements that are a predecessor
of an element in wavefront i + d.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-80-320.jpg)
![4.4. PARALLELISM 71
As demonstrated by O’Leary [163], SOR theory still holds in an approximate sense for multi-
colored matrices. The above observation that the Gauss-Seidel method with the natural ordering
is equivalent to a multicoloring cannot be extended to the SSOR method or wavefront-based in-
complete factorization preconditioners for the Conjugate Gradient method. In fact, tests by Duff
and Meurant [78] and an analysis by Eijkhout [84] show that multicolor incomplete factorization
preconditioners in general may take a considerably larger number of iterations to converge than
preconditioners based on the natural ordering. Whether this is offset by the increased parallelism
depends on the application and the computer architecture.
4.4.6 Blocked operations in the GMRES method
In addition to the usual matrix-vector product, inner products and vector updates, the precondi-
tioned GMRES method (see §2.3.4) has a kernel where one new vector, M−1
Av(j)
, is orthogonal-
ized against the previously built orthogonal set {v(1)
, v(2)
,..., v(j)
}. In our version, this is done
using Level 1 BLAS, which may be quite inefficient. To incorporate Level 2 BLAS we can apply
either Householder orthogonalization or classical Gram-Schmidt twice (which mitigates classical
Gram-Schmidt’s potential instability; see Saad [184]). Both approaches significantly increase the
computational work, but using classical Gram-Schmidt has the advantage that all inner products can
be performed simultaneously; that is, their communication can be packaged. This may increase the
efficiency of the computation significantly.
Another way to obtain more parallelism and data locality is to generate a basis {v(1)
, Av(1)
, ...,
Am
v(1)
} for the Krylov subspace first, and to orthogonalize this set afterwards; this is called m-step
GMRES(m) (see Kim and Chronopoulos [138]). (Compare this to the GMRES method in §2.3.4,
where each new vector is immediately orthogonalized to all previous vectors.) This approach
does not increase the computational work and, in contrast to CG, the numerical instability due to
generating a possibly near-dependent set is not necessarily a drawback.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-81-320.jpg)
![Chapter 5
Remaining topics
5.1 The Lanczos Connection
As discussed by Paige and Saunders in [167] and by Golub and Van Loan in [108], it is straightfor-
ward to derive the conjugate gradient method for solving symmetric positive definite linear systems
from the Lanczos algorithm for solving symmetric eigensystems and vice versa. As an example,
let us consider how one can derive the Lanczos process for symmetric eigensystems from the (un-
preconditioned) conjugate gradient method.
Suppose we define the n × k matrix R(k)
by
Rk = [r(0)
, r(1)
, . . . , r(k−1)
],
and the k × k upper bidiagonal matrix Bk by
Bk =
1 −β1 · · · 0
1 −β2
...
...
...
...
...
...
... −βk−1
0 · · · 1
,
where the sequences {r(k)
} and {βk} are defined by the standard conjugate gradient algorithm
discussed in §2.3.1. From the equations
p(j)
= r(j−1)
+ βj−1p(j−1)
, j = 2, 3, . . . , k ,
and p(1)
= r(0)
, we have Rk = PkBk, where
Pk = [p(1)
, p(2)
, . . . , p(k)
].
Assuming the elements of the sequence {p(j)
} are A-conjugate, it follows that
ˆTk = RT
k ARk = BT
k
ˆΛkBk
is a tridiagonal matrix since
ˆΛk =
p(1)T
Ap(1)
0 · · · 0
0 p(2)T
Ap(2)
...
...
...
...
...
...
... 0
0 · · · 0 p(k)T
Ap(k)
.
73](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-83-320.jpg)
![74 CHAPTER 5. REMAINING TOPICS
Since span{p(1)
, p(2)
, . . . , p(j)
} = span{r(0)
, r(1)
, . . . , r(j−1)
} and since the elements of
{r(j)
} are mutually orthogonal, it can be shown that the columns of n × k matrix Qk = Rk∆−1
form an orthonormal basis for the subspace span{b, Ab, . . . , Ak−1
b}, where ∆ is a diagonal
matrix whose ith diagonal element is r(i)
2. The columns of the matrix Qk are the Lanczos
vectors (see Parlett [170]) whose associated projection of A is the tridiagonal matrix
Tk = ∆−1
BT
k
ˆΛkBk∆−1
. (5.1)
The extremal eigenvalues of Tk approximate those of the matrix A. Hence, the diagonal and sub-
diagonal elements of Tk in (5.1), which are readily available during iterations of the conjugate
gradient algorithm (§2.3.1), can be used to construct Tk after k CG iterations. This allows us to
obtain good approximations to the extremal eigenvalues (and hence the condition number) of the
matrix A while we are generating approximations, x(k)
, to the solution of the linear system Ax = b.
For a nonsymmetric matrix A, an equivalent nonsymmetric Lanczos algorithm (see Lanc-
zos [141]) would produce a nonsymmetric matrix Tk in (5.1) whose extremal eigenvalues (which
may include complex-conjugate pairs) approximate those of A. The nonsymmetric Lanczos
method is equivalent to the BiCG method discussed in §2.3.5.
5.2 Block and s-step Iterative Methods
The methods discussed so far are all subspace methods, that is, in every iteration they extend the
dimension of the subspace generated. In fact, they generate an orthogonal basis for this subspace,
by orthogonalizing the newly generated vector with respect to the previous basis vectors.
However, in the case of nonsymmetric coefficient matrices the newly generated vector may be
almost linearly dependent on the existing basis. To prevent break-down or severe numerical error in
such instances, methods have been proposed that perform a look-ahead step (see Freund, Gutknecht
and Nachtigal [100], Parlett, Taylor and Liu [171], and Freund and Nachtigal [101]).
Several new, unorthogonalized, basis vectors are generated and are then orthogonalized with
respect to the subspace already generated. Instead of generating a basis, such a method generates a
series of low-dimensional orthogonal subspaces.
The s-step iterative methods of Chronopoulos and Gear [54] use this strategy of generating
unorthogonalized vectors and processing them as a block to reduce computational overhead and
improve processor cache behaviour.
If conjugate gradient methods are considered to generate a factorization of a tridiagonal re-
duction of the original matrix, then look-ahead methods generate a block factorization of a block
tridiagonal reduction of the matrix.
A block tridiagonal reduction is also effected by the Block Lanczos algorithm and the Block
Conjugate Gradient method (see O’Leary [162]). Such methods operate on multiple linear systems
with the same coefficient matrix simultaneously, for instance with multiple right hand sides, or the
same right hand side but with different initial guesses. Since these block methods use multiple
search directions in each step, their convergence behavior is better than for ordinary methods. In
fact, one can show that the spectrum of the matrix is effectively reduced by the nb − 1 smallest
eigenvalues, where nb is the block size.
5.3 Reduced System Preconditioning
As we have seen earlier, a suitable preconditioner for CG is a matrix M such that the system
M−1
Ax = M−1
f
requires fewer iterations to solve than Ax = f does, and for which systems Mz = r can be solved
efficiently. The first property is independent of the machine used, while the second is highly ma-
chine dependent. Choosing the best preconditioner involves balancing those two criteria in a way](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-84-320.jpg)
![5.4. DOMAIN DECOMPOSITION METHODS 75
that minimizes the overall computation time. One balancing approach used for matrices A arising
from 5-point finite difference discretization of second order elliptic partial differential equations
(PDEs) with Dirichlet boundary conditions involves solving a reduced system. Specifically, for an
n × n grid, we can use a point red-black ordering of the nodes to get
Ax =
DR C
CT
DB
xR
xB
=
fR
fB
, (5.2)
where DR and DB are diagonal, and C is a well-structured sparse matrix with 5 nonzero diagonals
if n is even and 4 nonzero diagonals if n is odd. Applying one step of block Gaussian elimination
(or computing the Schur complement; see Golub and Van Loan [108]) we have
I O
−CT
D−1
R I
DR C
CT
DB
xR
xB
=
I O
−CT
D−1
R I
fR
fB
which reduces to
DR C
O DB − CT
DR
−1
C
xR
xB
=
fR
fB − CT
DR
−1
fR
.
With proper scaling (left and right multiplication by DB
−1/2
), we obtain from the second block
equation the reduced system
(I − HT
H)y = g , (5.3)
where H = D
−1/2
R CD
−1/2
B , y = D
1/2
B xB, and g = D
−1/2
B (fB − CT
D−1
R fR). The linear system
(5.3) is of order n2
/2 for even n and of order (n2
− 1)/2 for odd n. Once y is determined, the
solution x is easily retrieved from y. The values on the black points are those that would be obtained
from a red/black ordered SSOR preconditioner on the full system, so we expect faster convergence.
The number of nonzero coefficients is reduced, although the coefficient matrix in (5.3) has
nine nonzero diagonals. Therefore it has higher density and offers more data locality. Meier and
Sameh [149] demonstrate that the reduced system approach on hierarchical memory machines such
as the Alliant FX/8 is over 3 times faster than unpreconditioned CG for Poisson’s equation on n×n
grids with n ≥ 250.
For 3-dimensional elliptic PDEs, the reduced system approach yields a block tridiagonal ma-
trix C in (5.2) having diagonal blocks of the structure of C from the 2-dimensional case and off-
diagonal blocks that are diagonal matrices. Computing the reduced system explicitly leads to an
unreasonable increase in the computational complexity of solving Ax = f. The matrix products
required to solve (5.3) would therefore be performed implicitly which could significantly decrease
performance. However, as Meier and Sameh show [149], the reduced system approach can still
be about 2-3 times as fast as the conjugate gradient method with Jacobi preconditioning for 3-
dimensional problems.
5.4 Domain Decomposition Methods
In recent years, much attention has been given to domain decomposition methods for linear el-
liptic problems that are based on a partitioning of the domain of the physical problem. Since the
subdomains can be handled independently, such methods are very attractive for coarse-grain par-
allel computers. On the other hand, it should be stressed that they can be very effective even on
sequential computers.
In this brief survey, we shall restrict ourselves to the standard second order self-adjoint scalar
elliptic problems in two dimensions of the form:
− · (a(x, y) u) = f(x, y) (5.4)](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-85-320.jpg)
![76 CHAPTER 5. REMAINING TOPICS
where a(x, y) is a positive function on the domain Ω, on whose boundary the value of u is pre-
scribed (the Dirichlet problem). For more general problems, and a good set of references, the reader
is referred to the series of proceedings [47, 48, 49, 106, 134, 176].
For the discretization of (5.4), we shall assume for simplicity that Ω is triangulated by a set TH
of nonoverlapping coarse triangles (subdomains) Ωi, i = 1, ..., p with nH internal vertices. The
Ωi’s are in turn further refined into a set of smaller triangles Th with n internal vertices in total.
Here H, h denote the coarse and fine mesh size respectively. By a standard Ritz-Galerkin method
using piecewise linear triangular basis elements on (5.4), we obtain an n × n symmetric positive
definite linear system Au = f.
Generally, there are two kinds of approaches depending on whether the subdomains overlap
with one another (Schwarz methods) or are separated from one another by interfaces (Schur Com-
plement methods, iterative substructuring).
We shall present domain decomposition methods as preconditioners for the linear system Au =
f or to a reduced (Schur Complement) system SBuB = gB defined on the interfaces in the non-
overlapping formulation. When used with the standard Krylov subspace methods discussed else-
where in this book, the user has to supply a procedure for computing Av or Sw given v or w and
the algorithms to be described herein computes K−1
v. The computation of Av is a simple sparse
matrix-vector multiply, but Sw may require subdomain solves, as will be described later.
5.4.1 Overlapping Subdomain Methods
In this approach, each substructure Ωi is extended to a larger substructure Ωi containing ni internal
vertices and all the triangles Ti ⊂ Th, within a distance δ from Ωi, where δ refers to the amount of
overlap.
Let Ai, AH denote the the discretizations of (5.4) on the subdomain triangulation Ti and the
coarse triangulation TH respectively.
Let RT
i denote the extension operator which extends (by zero) a function on Ti to Th and Ri the
corresponding pointwise restriction operator. Similarly, let RT
H denote the interpolation operator
which maps a function on the coarse grid TH onto the fine grid Th by piecewise linear interpolation
and RH the corresponding weighted restriction operator.
With these notations, the Additive Schwarz Preconditioner Kas for the system Au = f can be
compactly described as:
K−1
as v = RT
HA−1
H RHv +
p
i=1
RT
i A
−1
i Riv.
Note that the right hand side can be computed using p subdomain solves using the Ai’s, plus
a coarse grid solve using AH, all of which can be computed in parallel. Each term RT
i A
−1
i Riv
should be evaluated in three steps: (1) Restriction: vi = Riv, (2) Subdomain solves for wi: Aiwi =
vi, (3) Interpolation: yi = RT
i wi. The coarse grid solve is handled in the same manner.
The theory of Dryja and Widlund [75] shows that the condition number of K−1
as A is bounded
independently of both the coarse grid size H and the fine grid size h, provided there is “sufficient”
overlap between Ωi and Ωi (essentially it means that the ratio δ/H of the distance δ of the boundary
∂Ωi to ∂Ωi should be uniformly bounded from below as h → 0.) If the coarse grid solve term is left
out, then the condition number grows as O(1/H2
), reflecting the lack of global coupling provided
by the coarse grid.
For the purpose of implementations, it is often useful to interpret the definition of Kas in matrix
notation. Thus the decomposition of Ω into Ωi’s corresponds to partitioning of the components of
the vector u into p overlapping groups of index sets Ii’s, each with ni components. The ni × ni
matrix Ai is simply a principal submatrix of A corresponding to the index set Ii. RT
i is a n × ni
matrix defined by its action on a vector ui defined on Ti as: (RT
i ui)j = (ui)j if j ∈ Ii but
is zero otherwise. Similarly, the action of its transpose Riu forms an ni-vector by picking off](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-86-320.jpg)
![5.4. DOMAIN DECOMPOSITION METHODS 77
the components of u corresponding to Ii. Analogously, RT
H is an n × nH matrix with entries
corresponding to piecewise linear interpolation and its transpose can be interpreted as a weighted
restriction matrix. If Th is obtained from TH by nested refinement, the action of RH, RT
H can be
efficiently computed as in a standard multigrid algorithm. Note that the matrices RT
i , Ri, RT
H, RH
are defined by their actions and need not be stored explicitly.
We also note that in this algebraic formulation, the preconditioner Kas can be extended to
any matrix A, not necessarily one arising from a discretization of an elliptic problem. Once we
have the partitioning index sets Ii’s, the matrices Ri, Ai are defined. Furthermore, if A is positive
definite, then Ai is guaranteed to be nonsingular. The difficulty is in defining the “coarse grid”
matrices AH, RH, which inherently depends on knowledge of the grid structure. This part of the
preconditioner can be left out, at the expense of a deteriorating convergence rate as p increases.
Radicati and Robert [177] have experimented with such an algebraic overlapping block Jacobi
preconditioner.
5.4.2 Non-overlapping Subdomain Methods
The easiest way to describe this approach is through matrix notation. The set of vertices of Th
can be divided into two groups. The set of interior vertices I of all Ωi and the set of vertices B
which lies on the boundaries i ∂Ωi of the coarse triangles Ωi in TH. We shall re-order u and f as
u ≡ (uI, uB)T
and f ≡ (fI, fB)T
corresponding to this partition. In this ordering, equation (5.4)
can be written as follows:
AI AIB
AT
IB AB
uI
uB
=
fI
fB
. (5.5)
We note that since the subdomains are uncoupled by the boundary vertices, AI = blockdiagonal(Ai)
is block-diagonal with each block Ai being the stiffness matrix corresponding to the unknowns
belonging to the interior vertices of subdomain Ωi.
By a block LU-factorization of A, the system in (5.5) can be written as:
I 0
AT
IBA−1
I I
AI AIB
0 SB
uI
uB
=
fI
fB
, (5.6)
where
SB ≡ AB − AT
IBA−1
I AIB
is the Schur complement of AB in A.
By eliminating uI in (5.6), we arrive at the following equation for uB:
SBuB = gB ≡ fB − AIBA−1
I fI. (5.7)
We note the following properties of this Schur Complement system:
1. SB inherits the symmetric positive definiteness of A.
2. SB is dense in general and computing it explicitly requires as many solves on each subdo-
main as there are points on each of its edges.
3. The condition number of SB is O(h−1
), an improvement over the O(h−2
) growth for A.
4. Given a vector vB defined on the boundary vertices B of TH, the matrix-vector product
SBvB can be computed according to ABvB − AT
IB(A−1
I (AIBvB)) where A−1
I involves p
independent subdomain solves using A−1
i .
5. The right hand side gB can also be computed using p independent subdomain solves.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-87-320.jpg)
![78 CHAPTER 5. REMAINING TOPICS
These properties make it possible to apply a preconditioned iterative method to (5.7), which is
the basic method in the nonoverlapping substructuring approach. We will also need some good
preconditioners to further improve the convergence of the Schur system.
We shall first describe the Bramble-Pasciak-Schatz preconditioner [36]. For this, we need to
further decompose B into two non-overlapping index sets:
B = E ∪ VH (5.8)
where VH = k Vk denote the set of nodes corresponding to the vertices Vk’s of TH, and E =
i Ei denote the set of nodes on the edges Ei’s of the coarse triangles in TH, excluding the vertices
belonging to VH.
In addition to the coarse grid interpolation and restriction operators RH, RT
H defined before, we
shall also need a new set of interpolation and restriction operators for the edges Ei’s. Let REi
be
the pointwise restriction operator (an nEi × n matrix, where nEi is the number of vertices on the
edge Ei) onto the edge Ei defined by its action (REi uB)j = (uB)j if j ∈ Ei but is zero otherwise.
The action of its transpose extends by zero a function defined on Ei to one defined on B.
Corresponding to this partition of B, S can be written in the block form:
SB =
SE SEV
ST
EV SV
. (5.9)
The block SE can again be block partitioned, with most of the subblocks being zero. The diagonal
blocks SEi of SE are the principal submatrices of S corresponding to Ei. Each SEi represents the
coupling of nodes on interface Ei separating two neighboring subdomains.
In defining the preconditioner, the action of S−1
Ei
is needed. However, as noted before, in
general SEi
is a dense matrix which is also expensive to compute, and even if we had it, it would be
expensive to compute its action (we would need to compute its inverse or a Cholesky factorization).
Fortunately, many efficiently invertible approximations to SEi have been proposed in the literature
(see Keyes and Gropp [135]) and we shall use these so-called interface preconditioners for SEi
instead. We mention one specific preconditioner:
MEi
= αEi
K1/2
where K is an nEi × nEi one dimensional Laplacian matrix, namely a tridiagonal matrix with 2’s
down the main diagonal and −1’s down the two off-diagonals, and αEi is taken to be some average
of the coefficient a(x, y) of (5.4) on the edge Ei. We note that since the eigen-decomposition
of K is known and computable by the Fast Sine Transform, the action of MEi can be efficiently
computed.
With these notations, the Bramble-Pasciak-Schatz preconditioner is defined by its action on a
vector vB defined on B as follows:
K−1
BP SvB = RT
HA−1
H RHvB +
Ei
RT
Ei
M−1
Ei
REi vB. (5.10)
Analogous to the additive Schwarz preconditioner, the computation of each term consists of the
three steps of restriction-inversion-interpolation and is independent of the others and thus can be
carried out in parallel.
Bramble, Pasciak and Schatz [36] prove that the condition number of K−1
BP SSB is bounded by
O(1 + log2
(H/h)). In practice, there is a slight growth in the number of iterations as h becomes
small (i.e., as the fine grid is refined) or as H becomes large (i.e., as the coarse grid becomes
coarser).
The log2
(H/h) growth is due to the coupling of the unknowns on the edges incident on a com-
mon vertex Vk, which is not accounted for in KBP S. Smith [190] proposed a vertex space modifica-
tion to KBP S which explicitly accounts for this coupling and therefore eliminates the dependence
on H and h. The idea is to introduce further subsets of B called vertex spaces X = k Xk with](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-88-320.jpg)
![5.4. DOMAIN DECOMPOSITION METHODS 79
Xk consisting of a small set of vertices on the edges incident on the vertex Vk and adjacent to it.
Note that X overlaps with E and VH. Let SXk
be the principal submatrix of SB corresponding
to Xk, and RXk
, RT
Xk
be the corresponding restriction (pointwise) and extension (by zero) matri-
ces. As before, SXk
is dense and expensive to compute and factor/solve but efficiently invertable
approximations (some using variants of the K1/2
operator defined before) have been developed in
the literature (see Chan, Mathew and Shao [51]). We shall let MXk
be such a preconditioner for
SXk
. Then Smith’s Vertex Space preconditioner is defined by:
K−1
V SvB = RT
HA−1
H RHvB +
Ei
RT
Ei
M−1
Ei
REi vB
+
Xk
RT
Xk
M−1
Xk
RXk
vB. (5.11)
Smith [190] proved that the condition number of K−1
V SSB is bounded independent of H and h,
provided there is sufficient overlap of Xk with B.
5.4.3 Further Remarks
Multiplicative Schwarz Methods
As mentioned before, the Additive Schwarz preconditioner can be viewed as an overlapping block
Jacobi preconditioner. Analogously, one can define a multiplicative Schwarz preconditioner which
corresponds to a symmetric block Gauss-Seidel version. That is, the solves on each subdomain are
performed sequentially, using the most current iterates as boundary conditions from neighboring
subdomains. Even without conjugate gradient acceleration, the multiplicative method can take
many fewer iterations than the additive version. However, the multiplicative version is not as
parallelizable, although the degree of parallelism can be increased by trading off the convergence
rate through multi-coloring the subdomains. The theory can be found in Bramble, et al. [37].
Inexact Solves
The exact solves involving A
−1
i , A−1
i and A−1
H in Kas, KBP S, KV S can be replaced by inexact
solves ˜A
−1
i , ˜A−1
i and ˜A−1
H , which can be standard elliptic preconditioners themselves (e.g. multi-
grid, ILU, SSOR, etc.).
For the Schwarz methods, the modification is straightforward and the Inexact Solve Additive
Schwarz Preconditioner is simply:
˜K−1
as v = RT
H
˜A−1
H RHv +
p
i=1
RT
i
˜A
−1
i Riv.
The Schur Complement methods require more changes to accommodate inexact solves. By
replacing A−1
H by ˜A−1
H in the definitions of KBP S and KV S, we can easily obtain inexact pre-
conditioners ˜KBP S and ˜KV S for SB. The main difficulty is, however, that the evaluation of the
product SBvB requires exact subdomain solves in A−1
I . One way to get around this is to use an
inner iteration using ˜Ai as a preconditioner for Ai in order to compute the action of A−1
I . An
alternative is to perform the iteration on the larger system (5.5) and construct a preconditioner from
the factorization in (5.6) by replacing the terms AI, SB by ˜AI, ˜SB respectively, where ˜SB can be
either ˜KBP S or ˜KV S. Care must be taken to scale ˜AH and ˜Ai so that they are as close to AH and
Ai as possible respectively — it is not sufficient that the condition number of ˜A−1
H AH and ˜A−1
i Ai
be close to unity, because the scaling of the coupling matrix AIB may be wrong.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-89-320.jpg)
![80 CHAPTER 5. REMAINING TOPICS
Nonsymmetric Problems
The preconditioners given above extend naturally to nonsymmetric A’s (e.g., convection-diffusion
problems), at least when the nonsymmetric part is not too large. The nice theoretical convergence
rates can be retained provided that the coarse grid size H is chosen small enough (depending
on the size of the nonsymmetric part of A) (see Cai and Widlund [43]). Practical implementations
(especially for parallelism) of nonsymmetric domain decomposition methods are discussed in [136,
137].
Choice of Coarse Grid Size H
Given h, it has been observed empirically (see Gropp and Keyes [110]) that there often exists an
optimal value of H which minimizes the total computational time for solving the problem. A small
H provides a better, but more expensive, coarse grid approximation, and requires solving more, but
smaller, subdomain solves. A large H has the opposite effect. For model problems, the optimal H
can be determined for both sequential and parallel implementations (see Chan and Shao [52]). In
practice, it may pay to determine a near optimal value of H empirically if the preconditioner is to
be re-used many times. However, there may also be geometric constraints on the range of values
that H can take.
5.5 Multigrid Methods
Simple iterative methods (such as the Jacobi method) tend to damp out high frequency components
of the error fastest (see §2.2.1). This has led people to develop methods based on the following
heuristic:
1. Perform some steps of a basic method in order to smooth out the error.
2. Restrict the current state of the problem to a subset of the grid points, the so-called “coarse
grid”, and solve the resulting projected problem.
3. Interpolate the coarse grid solution back to the original grid, and perform a number of steps
of the basic method again.
Steps 1 and 3 are called “pre-smoothing” and “post-smoothing” respectively; by applying this
method recursively to step 2 it becomes a true “multigrid” method. Usually the generation of
subsequently coarser grids is halted at a point where the number of variables becomes small enough
that direct solution of the linear system is feasible.
The method outlined above is said to be a “V-cycle” method, since it descends through a se-
quence of subsequently coarser grids, and then ascends this sequence in reverse order. A “W-cycle”
method results from visiting the coarse grid twice, with possibly some smoothing steps in between.
An analysis of multigrid methods is relatively straightforward in the case of simple differential
operators such as the Poisson operator on tensor product grids. In that case, each next coarse grid
is taken to have the double grid spacing of the previous grid. In two dimensions, a coarse grid
will have one quarter of the number of points of the corresponding fine grid. Since the coarse grid
is again a tensor product grid, a Fourier analysis (see for instance Briggs [42]) can be used. For
the more general case of self-adjoint elliptic operators on arbitrary domains a more sophisticated
analysis is needed (see Hackbusch [116], McCormick [147]). Many multigrid methods can be
shown to have an (almost) optimal number of operations, that is, the work involved is proportional
to the number of variables.
From the above description it is clear that iterative methods play a role in multigrid theory as
smoothers (see Kettler [132]). Conversely, multigrid-like methods can be used as preconditioners](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-90-320.jpg)
![5.6. ROW PROJECTION METHODS 81
in iterative methods. The basic idea here is to partition the matrix on a given grid to a 2×2 structure
A(i)
=
A
(i)
1,1 A
(i)
1,2
A
(i)
2,1 A
(i)
2,2
with the variables in the second block row corresponding to the coarse grid nodes. The matrix on
the next grid is then an incomplete version of the Schur complement
A(i+1)
≈ S(i)
= A
(i)
2,2 − A
(i)
2,1A
(i)−1
1,1 A
(i)
1,2.
The coarse grid is typically formed based on a red-black or cyclic reduction ordering; see for
instance Rodrigue and Wolitzer [179], and Elman [92].
Some multigrid preconditioners try to obtain optimality results similar to those for the full
multigrid method. Here we will merely supply some pointers to the literature: Axelsson and Eijk-
hout [16], Axelsson and Vassilevski [23, 22], Braess [35], Maitre and Musy [144], McCormick and
Thomas [148], Yserentant [216] and Wesseling [213].
5.6 Row Projection Methods
Most iterative methods depend on spectral properties of the coefficient matrix, for instance some
require the eigenvalues to be in the right half plane. A class of methods without this limitation is
that of row projection methods (see Bj¨orck and Elfving [34], and Bramley and Sameh [38]). They
are based on a block row partitioning of the coefficient matrix
AT
= [A1, . . . , Am]
and iterative application of orthogonal projectors
Pix = Ai(AT
i Ai)−1
AT
i x.
These methods have good parallel properties and seem to be robust in handling nonsymmetric and
indefinite problems.
Row projection methods can be used as preconditioners in the conjugate gradient method. In
that case, there is a theoretical connection with the conjugate gradient method on the normal equa-
tions (see §2.3.3).](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-91-320.jpg)
![84 APPENDIX A. OBTAINING THE SOFTWARE
sh templates.shar
No subdirectory will be created. The unpacked files will stay in the current directory. Each
method is written as a separate subroutine in its own file, named after the method (e.g., CG.f,
BiCGSTAB.f, GMRES.f). The argument parameters are the same for each, with the exception of
the required matrix-vector products and preconditioner solvers (some require the transpose matrix).
Also, the amount of workspace needed varies. The details are documented in each routine.
Note that the vector-vector operations are accomplished using the BLAS [143] (many manu-
facturers have assembly coded these kernels for maximum performance), although a mask file is
provided to link to user defined routines.
The README file gives more details, along with instructions for a test routine.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-94-320.jpg)
![88 APPENDIX C. GLOSSARY
• Stalled: the measure of the error stays more or less constant during a number of iter-
ations. As above, this does not imply anything about the ultimate convergence of the
method.
Dense matrix Matrix for which the number of zero elements is too small to warrant specialized
algorithms to exploit these zeros.
Diagonally dominant matrix See: Matrix properties
Direct method An algorithm that produces the solution to a system of linear equations in a number
of operations that is determined a priori by the size of the system. In exact arithmetic, a direct
method yields the true solution to the system. See: Iterative method.
Distributed memory See: Parallel computer.
Divergence An iterative method is said to diverge if it does not converge in a reasonable number
of iterations, or if some measure of the error grows unacceptably. However, growth of the
error as such is no sign of divergence: a method with irregular convergence behavior may
ultimately converge, even though the error grows during some iterations.
Domain decomposition method Solution method for linear systems based on a partitioning of
the physical domain of the differential equation. Domain decomposition methods typically
involve (repeated) independent system solution on the subdomains, and some way of com-
bining data from the subdomains on the separator part of the domain.
Field of values Given a matrix A, the field of values is the set {xT
Ax : xT
x = 1}. For symmetric
matrices this is the range [λmin(A), λmax(A)].
Fill A position that is zero in the original matrix A but not in an exact factorization of A. In an
incomplete factorization, some fill elements are discarded.
Forward error The difference between a computed iterate and the true solution of a linear system,
measured in some vector norm.
Halfbandwidth See: Band matrix.
Ill-conditioned system A linear system for which the coefficient matrix has a large condition num-
ber. Since in many applications the condition number is proportional to (some power of) the
number of unknowns, this should be understood as the constant of proportionality being
large.
Incomplete factorization A factorization obtained by discarding certain elements during the fac-
torization process (‘modified’ and ‘relaxed’ incomplete factorization compensate in some
way for discarded elements). Thus an incomplete LU factorization of a matrix A will in
general satisfy A = LU; however, one hopes that the factorization LU will be close enough
to A to function as a preconditioner in an iterative method.
Iterate Approximation to the solution of a linear system in any iteration of an iterative method.
Iterative method An algorithm that produces a sequence of approximations to the solution of a
linear system of equations; the length of the sequence is not given a priori by the size of the
system. Usually, the longer one iterates, the closer one is able to get to the true solution. See:
Direct method.
Krylov sequence For a given matrix A and vector x, the sequence of vectors {Ai
x}i≥0, or a finite
initial part of this sequence.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-98-320.jpg)
![Bibliography
[1] J. AARDEN AND K.-E. KARLSSON, Preconditioned CG-type methods for solving the cou-
pled systems of fundamental semiconductor equations, BIT, 29 (1989), pp. 916–937.
[2] L. ADAMS AND H. JORDAN, Is SOR color-blind?, SIAM J. Sci. Statist. Comput., 7 (1986),
pp. 490–506.
[3] E. ANDERSON, ET. AL., LAPACK Users Guide, SIAM, Philadelphia, 1992.
[4] J. APPLEYARD AND I. CHESHIRE, Nested factorization, in Reservoir Simulation Sympo-
sium of the SPE, 1983. Paper 12264.
[5] M. ARIOLI, J. DEMMEL, AND I. DUFF, Solving sparse linear systems with sparse backward
error, SIAM J. Matrix Anal. Appl., 10 (1989), pp. 165–190.
[6] W. ARNOLDI, The principle of minimized iterations in the solution of the matrix eigenvalue
problem, Quart. Appl. Math., 9 (1951), pp. 17–29.
[7] S. ASHBY, CHEBYCODE: A Fortran implementation of Manteuffel’s adaptive Chebyshev
algorithm, Tech. Rep. UIUCDCS-R-85-1203, University of Illinois, 1985.
[8] S. ASHBY, T. MANTEUFFEL, AND J. OTTO, A comparison of adaptive Chebyshev and least
squares polynomial preconditioning for Hermitian positive definite linear systems, SIAM J.
Sci. Statist. Comput., 13 (1992), pp. 1–29.
[9] S. ASHBY, T. MANTEUFFEL, AND P. SAYLOR, Adaptive polynomial preconditioning for
Hermitian indefinite linear systems, BIT, 29 (1989), pp. 583–609.
[10] S. F. ASHBY, T. A. MANTEUFFEL, AND P. E. SAYLOR, A taxonomy for conjugate gradient
methods, SIAM J. Numer. Anal., 27 (1990), pp. 1542–1568.
[11] C. ASHCRAFT AND R. GRIMES, On vectorizing incomplete factorizations and SSOR pre-
conditioners, SIAM J. Sci. Statist. Comput., 9 (1988), pp. 122–151.
[12] O. AXELSSON, Incomplete block matrix factorization preconditioning methods. The ulti-
mate answer?, J. Comput. Appl. Math., 12&13 (1985), pp. 3–18.
[13] , A general incomplete block-matrix factorization method, Linear Algebra Appl., 74
(1986), pp. 179–190.
[14] O. AXELSSON AND A. BARKER, Finite element solution of boundary value problems. The-
ory and computation, Academic Press, Orlando, Fl., 1984.
[15] O. AXELSSON AND V. EIJKHOUT, Vectorizable preconditioners for elliptic difference equa-
tions in three space dimensions, J. Comput. Appl. Math., 27 (1989), pp. 299–321.
93](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-103-320.jpg)
![94 BIBLIOGRAPHY
[16] , The nested recursive two-level factorization method for nine-point difference matri-
ces, SIAM J. Sci. Statist. Comput., 12 (1991), pp. 1373–1400.
[17] O. AXELSSON AND I. GUSTAFSSON, Iterative solution for the solution of the Navier equa-
tions of elasticity, Comput. Methods Appl. Mech. Engrg., 15 (1978), pp. 241–258.
[18] O. AXELSSON AND G. LINDSKOG, On the eigenvalue distribution of a class of precondi-
tioning matrices, Numer. Math., 48 (1986), pp. 479–498.
[19] , On the rate of convergence of the preconditioned conjugate gradient method, Numer.
Math., 48 (1986), pp. 499–523.
[20] O. AXELSSON AND N. MUNKSGAARD, Analysis of incomplete factorizations with fixed
storage allocation, in Preconditioning Methods – Theory and Applications, D. Evans, ed.,
Gordon and Breach, New York, 1983, pp. 265–293.
[21] O. AXELSSON AND B. POLMAN, On approximate factorization methods for block-matrices
suitable for vector and parallel processors, Linear Algebra Appl., 77 (1986), pp. 3–26.
[22] O. AXELSSON AND P. VASSILEVSKI, Algebraic multilevel preconditioning methods, I, Nu-
mer. Math., 56 (1989), pp. 157–177.
[23] , Algebraic multilevel preconditioning methods, II, SIAM J. Numer. Anal., 57 (1990),
pp. 1569–1590.
[24] O. AXELSSON AND P. S. VASSILEVSKI, A black box generalized conjugate gradient solver
with inner iterations and variable-step preconditioning, SIAM J. Matrix Anal. Appl., 12
(1991), pp. 625–644.
[25] R. BANK, Marching algorithms for elliptic boundary value problems; II: The variable co-
efficient case, SIAM J. Numer. Anal., 14 (1977), pp. 950–970.
[26] R. BANK, T. CHAN, W. COUGHRAN JR., AND R. SMITH, The Alternate-Block-
Factorization procedure for systems of partial differential equations, BIT, 29 (1989),
pp. 938–954.
[27] R. BANK AND D. ROSE, Marching algorithms for elliptic boundary value problems. I: The
constant coefficient case, SIAM J. Numer. Anal., 14 (1977), pp. 792–829.
[28] R. E. BANK AND T. F. CHAN, A composite step bi-conjugate gradient algorithm for non-
symmetric linear systems, Tech. Rep. CAM 92-, UCLA, Dept. of Math., Los Angeles, CA
90024-1555, 1992.
[29] R. E. BANK AND T. F. CHAN, An analysis of the composite step Biconjugate gradient
method, Numerische Mathematik, 66 (1993), pp. 295–319.
[30] G. BAUDET, Asynchronous iterative methods for multiprocessors, J. Assoc. Comput. Mach.,
25 (1978), pp. 226–244.
[31] R. BEAUWENS, On Axelsson’s perturbations, Linear Algebra Appl., 68 (1985), pp. 221–
242.
[32] , Approximate factorizations with S/P consistently ordered M-factors, BIT, 29 (1989),
pp. 658–681.
[33] R. BEAUWENS AND L. QUENON, Existence criteria for partial matrix factorizations in
iterative methods, SIAM J. Numer. Anal., 13 (1976), pp. 615–643.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-104-320.jpg)
![BIBLIOGRAPHY 95
[34] A. BJ ¨ORCK AND T. ELFVING, Accelerated projection methods for computing pseudo-
inverse solutions of systems of linear equations, BIT, 19 (1979), pp. 145–163.
[35] D. BRAESS, The contraction number of a multigrid method for solving the Poisson equation,
Numer. Math., 37 (1981), pp. 387–404.
[36] J. H. BRAMBLE, J. E. PASCIAK, AND A. H. SCHATZ, The construction of preconditioners
for elliptic problems by substructuring, I, Mathematics of Computation, 47 (1986), pp. 103–
134.
[37] J. H. BRAMBLE, J. E. PASCIAK, J. WANG, AND J. XU, Convergence estimates for product
iterative methods with applications to domain decompositions and multigrid, Math. Comp.,
to appear.
[38] R. BRAMLEY AND A. SAMEH, Row projection methods for large nonsymmetric linear sys-
tems, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 168–193.
[39] C. BREZINSKI AND H. SADOK, Avoiding breakdown in the CGS algorithm, Numer. Alg.,
1 (1991), pp. 199–206.
[40] C. BREZINSKI, M. ZAGLIA, AND H. SADOK, Avoiding breakdown and near breakdown in
Lanczos type algorithms, Numer. Alg., 1 (1991), pp. 261–284.
[41] , A breakdown free Lanczos type algorithm for solving linear systems, Numer. Math.,
63 (1992), pp. 29–38.
[42] W. BRIGGS, A Multigrid Tutorial, SIAM, Philadelphia, 1977.
[43] X.-C. CAI AND O. WIDLUND, Multiplicative Schwarz algorithms for some nonsymmetric
and indefinite problems, SIAM J. Numer. Anal., 30 (1993), pp. 936–952.
[44] T. CHAN, Fourier analysis of relaxed incomplete factorization preconditioners, SIAM J.
Sci. Statist. Comput., 12 (1991), pp. 668–680.
[45] T. CHAN, L. DE PILLIS, AND H. VAN DER VORST, A transpose-free squared Lanczos
algorithm and application to solving nonsymmetric linear systems, Tech. Rep. CAM 91-17,
UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1991.
[46] T. CHAN, E. GALLOPOULOS, V. SIMONCINI, T. SZETO, AND C. TONG, A quasi-minimal
residual variant of the Bi-CGSTAB algorithm for nonsymmetric systems, Tech. Rep. CAM
92-26, UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1992. SIAM J. Sci. Comput.,
to appear.
[47] T. CHAN, R. GLOWINSKI, , J. P´ERIAUX, AND O. WIDLUND, eds., Domain Decomposition
Methods, Philadelphia, 1989, SIAM. Proceedings of the Second International Symposium
on Domain Decomposition Methods, Los Angeles, CA, January 14 - 16, 1988.
[48] , eds., Domain Decomposition Methods, Philadelphia, 1990, SIAM. Proceedings of the
Third International Symposium on Domain Decomposition Methods, Houston, TX, 1989.
[49] , eds., Domain Decomposition Methods, SIAM, Philadelphia, 1991. Proceedings of
the Fourth International Symposium on Domain Decomposition Methods, Moscow, USSR,
1990.
[50] T. CHAN AND C.-C. J. KUO, Two-color Fourier analysis of iterative algorithms for elliptic
problems with red/black ordering, SIAM J. Sci. Statist. Comput., 11 (1990), pp. 767–793.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-105-320.jpg)
![96 BIBLIOGRAPHY
[51] T. F. CHAN, T. P. MATHEW, AND J. P. SHAO, Efficient variants of the vertex space domain
decomposition algorithm, Tech. Rep. CAM 92-07, UCLA, Dept. of Math., Los Angeles, CA
90024-1555, 1992. SIAM J. Sci. Comput., to appear.
[52] T. F. CHAN AND J. SHAO, On the choice of coarse grid size in domain decomposition
methods, tech. rep., UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1993. to appear.
[53] D. CHAZAN AND W. MIRANKER, Chaotic relaxation, Linear Algebra Appl., 2 (1969),
pp. 199–222.
[54] A. CHRONOPOULOS AND C. GEAR, s-step iterative methods for symmetric linear systems,
J. Comput. Appl. Math., 25 (1989), pp. 153–168.
[55] P. CONCUS AND G. GOLUB, A generalized conjugate gradient method for nonsymmet-
ric systems of linear equations, in Computer methods in Applied Sciences and Engineering,
Second International Symposium, Dec 15–19, 1975; Lecture Notes in Economics and Math-
ematical Systems, Vol. 134, Berlin, New York, 1976, Springer-Verlag.
[56] P. CONCUS, G. GOLUB, AND G. MEURANT, Block preconditioning for the conjugate gra-
dient method, SIAM J. Sci. Statist. Comput., 6 (1985), pp. 220–252.
[57] P. CONCUS, G. GOLUB, AND D. O’LEARY, A generalized conjugate gradient method for
the numerical solution of elliptic partial differential equations, in Sparse Matrix Computa-
tions, J. Bunch and D. Rose, eds., Academic Press, New York, 1976, pp. 309–332.
[58] P. CONCUS AND G. H. GOLUB, Use of fast direct methods for the efficient numerical solu-
tion of nonseparable elliptic equations, SIAM J. Numer. Anal., 10 (1973), pp. 1103–1120.
[59] E. CUTHILL AND J. MCKEE, Reducing the bandwidth of sparse symmetric matrices, in
ACM Proceedings of the 24th National Conference, 1969.
[60] E. D’AZEVEDO, V. EIJKHOUT, AND C. ROMINE, LAPACK working note 56: Reducing
communication costs in the conjugate gradient algorithm on distributed memory multipro-
cessor, tech. report, Computer Science Department, University of Tennessee, Knoxville,
TN, 1993.
[61] E. D’AZEVEDO AND C. ROMINE, Reducing communication costs in the conjugate gradient
algorithm on distributed memory multiprocessors, Tech. Rep. ORNL/TM-12192, Oak Ridge
National Lab, Oak Ridge, TN, 1992.
[62] E. DE STURLER, A parallel restructured version of GMRES(m), Tech. Rep. 91-85, Delft
University of Technology, Delft, The Netherlands, 1991.
[63] E. DE STURLER AND D. R. FOKKEMA, Nested Krylov methods and preserving the orthog-
onality, Tech. Rep. Preprint 796, Utrecht University, Utrecht, The Netherlands, 1993.
[64] S. DEMKO, W. MOSS, AND P. SMITH, Decay rates for inverses of band matrices, Mathe-
matics of Computation, 43 (1984), pp. 491–499.
[65] J. DEMMEL, The condition number of equivalence transformations that block diagonalize
matrix pencils, SIAM J. Numer. Anal., 20 (1983), pp. 599–610.
[66] J. DEMMEL, M. HEATH, AND H. VAN DER VORST, Parallel linear algebra, in Acta Nu-
merica, Vol. 2, Cambridge Press, New York, 1993.
[67] S. DOI, On parallelism and convergence of incomplete LU factorizations, Appl. Numer.
Math., 7 (1991), pp. 417–436.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-106-320.jpg)
![BIBLIOGRAPHY 97
[68] J. DONGARRA, J. DUCROZ, I. DUFF, AND S. HAMMARLING, A set of level 3 Basic Linear
Algebra Subprograms, ACM Trans. Math. Soft., 16 (1990), pp. 1–17.
[69] J. DONGARRA, J. DUCROZ, S. HAMMARLING, AND R. HANSON, An extended set of
FORTRAN Basic Linear Algebra Subprograms, ACM Trans. Math. Soft., 14 (1988), pp. 1–
32.
[70] J. DONGARRA, I. DUFF, D. SORENSEN, AND H. VAN DER VORST, Solving Linear Systems
on Vector and Shared Memory Computers, SIAM, Philadelphia, PA, 1991.
[71] J. DONGARRA AND E. GROSSE, Distribution of mathematical software via electronic mail,
Comm. ACM, 30 (1987), pp. 403–407.
[72] J. DONGARRA, C. MOLER, J. BUNCH, AND G. STEWART, LINPACK Users’ Guide, SIAM,
Philadelphia, 1979.
[73] J. DONGARRA AND H. VAN DER VORST, Performance of various computers using stan-
dard sparse linear equations solving techniques, in Computer Benchmarks, J. Dongarra and
W. Gentzsch, eds., Elsevier Science Publishers B.V., New York, 1993, pp. 177–188.
[74] F. DORR, The direct solution of the discrete Poisson equation on a rectangle, SIAM Rev.,
12 (1970), pp. 248–263.
[75] M. DRYJA AND O. B. WIDLUND, Towards a unified theory of domain decomposition algo-
rithms for elliptic problems, Tech. Rep. 486, also Ultracomputer Note 167, Department of
Computer Science, Courant Institute, 1989.
[76] D. DUBOIS, A. GREENBAUM, AND G. RODRIGUE, Approximating the inverse of a matrix
for use in iterative algorithms on vector processors, Computing, 22 (1979), pp. 257–268.
[77] I. DUFF, R. GRIMES, AND J. LEWIS, Sparse matrix test problems, ACM Trans. Math. Soft.,
15 (1989), pp. 1–14.
[78] I. DUFF AND G. MEURANT, The effect of ordering on preconditioned conjugate gradients,
BIT, 29 (1989), pp. 635–657.
[79] I. S. DUFF, A. M. ERISMAN, AND J.K.REID, Direct methods for sparse matrices, Oxford
University Press, London, 1986.
[80] T. DUPONT, R. KENDALL, AND H. RACHFORD, An approximate factorization proce-
dure for solving self-adjoint elliptic difference equations, SIAM J. Numer. Anal., 5 (1968),
pp. 559–573.
[81] E. D’YAKONOV, The method of variable directions in solving systems of finite difference
equations, Soviet Math. Dokl., 2 (1961), pp. 577–580. TOM 138, 271–274.
[82] L. EHRLICH, An Ad-Hoc SOR method, J. Comput. Phys., 43 (1981), pp. 31–45.
[83] M. EIERMANN AND R. VARGA, Is the optimal ω best for the SOR iteration method?, Linear
Algebra Appl., 182 (1993), pp. 257–277.
[84] V. EIJKHOUT, Analysis of parallel incomplete point factorizations, Linear Algebra Appl.,
154–156 (1991), pp. 723–740.
[85] , Beware of unperturbed modified incomplete point factorizations, in Proceedings of
the IMACS International Symposium on Iterative Methods in Linear Algebra, Brussels, Bel-
gium, R. Beauwens and P. de Groen, eds., 1992.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-107-320.jpg)
![98 BIBLIOGRAPHY
[86] , LAPACK working note 50: Distributed sparse data structures for linear algebra
operations, Tech. Rep. CS 92-169, Computer Science Department, University of Tennessee,
Knoxville, TN, 1992.
[87] , LAPACK working note 51: Qualitative properties of the conjugate gradient and Lanc-
zos methods in a matrix framework, Tech. Rep. CS 92-170, Computer Science Department,
University of Tennessee, Knoxville, TN, 1992.
[88] V. EIJKHOUT AND B. POLMAN, Decay rates of inverses of banded M-matrices that are
near to Toeplitz matrices, Linear Algebra Appl., 109 (1988), pp. 247–277.
[89] V. EIJKHOUT AND P. VASSILEVSKI, Positive definitess aspects of vectorizable precondi-
tioners, Parallel Computing, 10 (1989), pp. 93–100.
[90] S. EISENSTAT, Efficient implementation of a class of preconditioned conjugate gradient
methods, SIAM J. Sci. Statist. Comput., 2 (1981), pp. 1–4.
[91] R. ELKIN, Convergence theorems for Gauss-Seidel and other minimization algorithms,
Tech. Rep. 68-59, Computer Science Center, University of Maryland, College Park, MD,
Jan. 1968.
[92] H. ELMAN, Approximate Schur complement preconditioners on serial and parallel comput-
ers, SIAM J. Sci. Statist. Comput., 10 (1989), pp. 581–605.
[93] H. ELMAN AND M. SCHULTZ, Preconditioning by fast direct methods for non self-adjoint
nonseparable elliptic equations, SIAM J. Numer. Anal., 23 (1986), pp. 44–57.
[94] L. ELSNER, A note on optimal block-scaling of matrices, Numer. Math., 44 (1984), pp. 127–
128.
[95] V. FABER AND T. MANTEUFFEL, Necessary and sufficient conditions for the existence of a
conjugate gradient method, SIAM J. Numer. Anal., 21 (1984), pp. 315–339.
[96] G. FAIRWEATHER, A. GOURLAY, AND A. MITCHELL, Some high accuracy difference
schemes with a splitting operator for equations of parabolic and elliptic type, Numer. Math.,
10 (1967), pp. 56–66.
[97] R. FLETCHER, Conjugate gradient methods for indefinite systems, in Numerical Analysis
Dundee 1975, G. Watson, ed., Berlin, New York, 1976, Springer Verlag, pp. 73–89.
[98] G. FORSYTHE AND E. STRAUSS, On best conditioned matrices, Proc. Amer. Math. Soc., 6
(1955), pp. 340–345.
[99] R. FREUND, Conjugate gradient-type methods for linear systems with complex symmetric
coefficient matrices, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 425–448.
[100] R. FREUND, M. GUTKNECHT, AND N. NACHTIGAL, An implementation of the look-ahead
Lanczos algorithm for non-Hermitian matrices, SIAM J. Sci. Comput., 14 (1993), pp. 137–
158.
[101] R. FREUND AND N. NACHTIGAL, QMR: A quasi-minimal residual method for non-
Hermitian linear systems, Numer. Math., 60 (1991), pp. 315–339.
[102] , An implementation of the QMR method based on coupled two-term recurrences, Tech.
Rep. 92.15, RIACS, NASA Ames, Ames, CA, 1992.
[103] R. FREUND AND T. SZETO, A quasi-minimal residual squared algorithm for non-Hermitian
linear systems, tech. rep., RIACS, NASA Ames, Ames, CA, 1991.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-108-320.jpg)
![BIBLIOGRAPHY 99
[104] R. W. FREUND, A transpose-free quasi-minimum residual algorithm for non-Hermitian lin-
ear systems, SIAM J. Sci. Comput., 14 (1993), pp. 470–482.
[105] R. W. FREUND, G. H. GOLUB, AND N. M. NACHTIGAL, Iterative solution of linear sys-
tems, Acta Numerica, (1992), pp. 57–100.
[106] R. GLOWINSKI, G. H. GOLUB, G. A. MEURANT, AND J. P´ERIAUX, eds., Domain Decom-
position Methods for Partial Differential Equations, SIAM, Philadelphia, 1988. Proceedings
of the First International Symposium on Domain Decomposition Methods for Partial Differ-
ential Equations, Paris, France, January 1987.
[107] G. GOLUB AND D. O’LEARY, Some history of the conjugate gradient and Lanczos meth-
ods, SIAM Rev., 31 (1989), pp. 50–102.
[108] G. GOLUB AND C. VAN LOAN, Matrix Computations, second edition, The Johns Hopkins
University Press, Baltimore, 1989.
[109] A. GREENBAUM AND Z. STRAKOS, Predicting the behavior of finite precision Lanczos and
conjugate gradient computations, SIAM J. Mat. Anal. Appl., 13 (1992), pp. 121–137.
[110] W. D. GROPP AND D. E. KEYES, Domain decomposition with local mesh refinement, SIAM
J. Sci. Statist. Comput., 13 (1992), pp. 967–993.
[111] I. GUSTAFSSON, A class of first-order factorization methods, BIT, 18 (1978), pp. 142–156.
[112] M. H. GUTKNECHT, A completed theory of the unsymmetric Lanczos process and related
algorithms, part II, Tech. Rep. 90-16, IPS Research Report, ETH Z¨urich, Switzerland, 1990.
[113] , The unsymmetric Lanczos algorithms and their relations to P´ade approximation, con-
tinued fractions and the QD algorithm, in Proceedings of the Copper Mountain Conference
on Iterative Methods, 1990.
[114] , Variants of Bi-CGSTAB for matrices with complex spectrum, Tech. Rep. 91-14, IPS
ETH, Z¨urich, Switzerland, 1991.
[115] , A completed theory of the unsymmetric Lanczos process and related algorithms, part
I, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 594–639.
[116] W. HACKBUSCH, Multi-Grid Methods and Applications, Springer-Verlag, Berlin, New
York, 1985.
[117] , Iterative L¨osung großer schwachbesetzter Gleichungssysteme, Teubner, Stuttgart,
1991.
[118] A. HADJIDIMOS, On some high accuracy difference schemes for solving elliptic equations,
Numer. Math., 13 (1969), pp. 396–403.
[119] L. HAGEMAN AND D. YOUNG, Applied Iterative Methods, Academic Press, New York,
1981.
[120] W. HAGER, Condition estimators, SIAM J. Sci. Statist. Comput., 5 (1984), pp. 311–316.
[121] M. HESTENES AND E. STIEFEL, Methods of conjugate gradients for solving linear systems,
J. Res. Nat. Bur. Stand., 49 (1952), pp. 409–436.
[122] M. R. HESTENES, Conjugacy and gradients, in A History of Scientific Computing,
Addison-Wesley, Reading, MA, 1990, pp. 167–179.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-109-320.jpg)
![100 BIBLIOGRAPHY
[123] N. HIGHAM, Experience with a matrix norm estimator, SIAM J. Sci. Statist. Comput., 11
(1990), pp. 804–809.
[124] K. JEA AND D. YOUNG, Generalized conjugate-gradient acceleration of nonsym- metriz-
able iterative methods, Linear Algebra Appl., 34 (1980), pp. 159–194.
[125] O. JOHNSON, C. MICCHELLI, AND G. PAUL, Polynomial preconditioning for conjugate
gradient calculation, SIAM J. Numer. Anal., 20 (1983), pp. 362–376.
[126] M. JONES AND P. PLASSMANN, Parallel solution of unstructed, sparse systems of linear
equations, in Proceedings of the Sixth SIAM conference on Parallel Processing for Scien-
tific Computing, R. Sincovec, D. Keyes, M. Leuze, L. Petzold, and D. Reed, eds., SIAM,
Philadelphia, pp. 471–475.
[127] , A parallel graph coloring heuristic, SIAM J. Sci. Statist. Comput., 14 (1993),
pp. 654–669.
[128] W. JOUBERT, Lanczos methods for the solution of nonsymmetric systems of linear equations,
SIAM J. Matrix Anal. Appl., 13 (1992), pp. 926–943.
[129] W. KAHAN, Gauss-Seidel methods of solving large systems of linear equations, PhD thesis,
University of Toronto, 1958.
[130] S. KANIEL, Estimates for some computational techniques in linear algebra, Mathematics
of Computation, 20 (1966), pp. 369–378.
[131] D. KERSHAW, The incomplete Cholesky-conjugate gradient method for the iterative solu-
tion of systems of linear equations, J. Comput. Phys., 26 (1978), pp. 43–65.
[132] R. KETTLER, Analysis and comparison of relaxation schemes in robust multigrid and pre-
conditioned conjugate gradient methods, in Multigrid Methods, Lecture Notes in Mathemat-
ics 960, W. Hackbusch and U. Trottenberg, eds., Springer-Verlag, Berlin, New York, 1982,
pp. 502–534.
[133] , Linear multigrid methods in numerical reservoir simulation, PhD thesis, Delft Uni-
versity of Technology, Delft, The Netherlands, 1987.
[134] D. E. KEYES, T. F. CHAN, G. MEURANT, J. S. SCROGGS, AND R. G. VOIGT, eds.,
Domain Decomposition Methods For Partial Differential Equations, SIAM, Philadelphia,
1992. Proceedings of the Fifth International Symposium on Domain Decomposition Meth-
ods, Norfolk, VA, 1991.
[135] D. E. KEYES AND W. D. GROPP, A comparison of domain decomposition techniques for
elliptic partial differential equations and their parallel implementation, SIAM J. Sci. Statist.
Comput., 8 (1987), pp. s166 – s202.
[136] , Domain decomposition for nonsymmetric systems of equations: Examples from com-
putational fluid dynamics, in Domain Decomposition Methods, proceedings of the Sec-
ond Internation Symposium, Los Angeles, California, January 14–16, 1988, T. F. Chan,
R. Glowinski, J. Periaux, and O. B. Widlund, eds., Philadelphia, 1989, SIAM, pp. 373–384.
[137] , Domain decomposition techniques for the parallel solution of nonsymmetric systems
of elliptic boundary value problems, Applied Num. Math., 6 (1989/1990), pp. 281–301.
[138] S. K. KIM AND A. T. CHRONOPOULOS, A class of Lanczos-like algorithms implemented
on parallel computers, Parallel Comput., 17 (1991), pp. 763–778.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-110-320.jpg)
![BIBLIOGRAPHY 101
[139] D. R. KINCAID, J. R. RESPESS, D. M. YOUNG, AND R. G. GRIMES, ITPACK 2C: A
Fortran package for solving large sparse linear systems by adaptive accelerated iterative
methods, ACM Trans. Math. Soft., 8 (1982), pp. 302–322. Algorithm 586.
[140] L. Y. KOLOTILINA AND A. Y. YEREMIN, On a family of two-level preconditionings of the
incomlete block factorization type, Sov. J. Numer. Anal. Math. Modelling, (1986), pp. 293–
320.
[141] C. LANCZOS, An iteration method for the solution of the eigenvalue problem of linear dif-
ferential and integral operators, J. Res. Nat. Bur. Stand., 45 (1950), pp. 255–282.
[142] , Solution of systems of linear equations by minimized iterations, J. Res. Nat. Bur.
Stand., 49 (1952), pp. 33–53.
[143] C. LAWSON, R. HANSON, D. KINCAID, AND F. KROGH, Basic Linear Algebra Subpro-
grams for FORTRAN usage, ACM Trans. Math. Soft., 5 (1979), pp. 308–325.
[144] J. MAITRE AND F. MUSY, The contraction number of a class of two-level methods; an exact
evaluation for some finite element subspaces and model problems, in Multigrid methods,
Proceedings, K¨oln-Porz, 1981, W. Hackbusch and U. Trottenberg, eds., vol. 960 of Lecture
Notes in Mathematics, 1982, pp. 535–544.
[145] T. MANTEUFFEL, The Tchebychev iteration for nonsymmetric linear systems, Numer. Math.,
28 (1977), pp. 307–327.
[146] , An incomplete factorization technique for positive definite linear systems, Mathemat-
ics of Computation, 34 (1980), pp. 473–497.
[147] S. MCCORMICK, Multilevel Adaptive Methods for Partial Differential Equations, SIAM,
Philadelphia, 1989.
[148] S. MCCORMICK AND J. THOMAS, The Fast Adaptive Composite grid (FAC) method for
elliptic equations, Mathematics of Computation, 46 (1986), pp. 439–456.
[149] U. MEIER AND A. SAMEH, The behavior of conjugate gradient algorithms on a multivector
processor with a hierarchical memory, J. Comput. Appl. Math., 24 (1988), pp. 13–32.
[150] U. MEIER-YANG, Preconditioned conjugate gradient-like methods for nonsymmetric linear
systems, tech. rep., CSRD, University of Illinois, Urbana, IL, April 1992.
[151] J. MEIJERINK AND H. VAN DER VORST, An iterative solution method for linear systems
of which the coefficient matrix is a symmetric M-matrix, Mathematics of Computation, 31
(1977), pp. 148–162.
[152] , Guidelines for the usage of incomplete decompositions in solving sets of linear equa-
tions as they occur in practical problems, J. Comput. Phys., 44 (1981), pp. 134–155.
[153] R. MELHEM, Toward efficient implementation of preconditioned conjugate gradient meth-
ods on vector supercomputers, Internat. J. Supercomput. Appls., 1 (1987), pp. 77–98.
[154] G. MEURANT, The block preconditioned conjugate gradient method on vector computers,
BIT, 24 (1984), pp. 623–633.
[155] , Multitasking the conjugate gradient method on the CRAY X-MP/48, Parallel Comput.,
5 (1987), pp. 267–280.
[156] N. MUNKSGAARD, Solving sparse symmetric sets of linear equations by preconditioned
conjugate gradients, ACM Trans. Math. Software, 6 (1980), pp. 206–219.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-111-320.jpg)
![102 BIBLIOGRAPHY
[157] N. NACHTIGAL, S. REDDY, AND L. TREFETHEN, How fast are nonsymmetric matrix iter-
ations?, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 778–795.
[158] N. NACHTIGAL, L. REICHEL, AND L. TREFETHEN, A hybrid GMRES algorithm for non-
symmetric matrix iterations, Tech. Rep. 90-7, MIT, Cambridge, MA, 1990.
[159] N. M. NACHTIGAL, A Look-Ahead Variant of the Lanczos Algorithm and its Application to
the Quasi-Minimal Residual Methods for Non-Hermitian Linear Systems, PhD thesis, MIT,
Cambridge, MA, 1991.
[160] Y. NOTAY, Solving positive (semi)definite linear systems by preconditioned iterative meth-
ods, in Preconditioned Conjugate Gradient Methods, O. Axelsson and L. Kolotilina, eds.,
vol. 1457 of Lecture Notes in Mathematics, Nijmegen, 1989, pp. 105–125.
[161] , On the robustness of modified incomplete factorization methods, Internat. J. Comput.
Math., 40 (1992), pp. 121–141.
[162] D. O’LEARY, The block conjugate gradient algorithm and related methods, Linear Algebra
Appl., 29 (1980), pp. 293–322.
[163] , Ordering schemes for parallel processing of certain mesh problems, SIAM J. Sci.
Statist. Comput., 5 (1984), pp. 620–632.
[164] T. C. OPPE, W. D. JOUBERT, AND D. R. KINCAID, NSPCG user’s guide, version 1.0: A
package for solving large sparse linear systems by various iterative methods, Tech. Rep.
CNA–216, Center for Numerical Analysis, University of Texas at Austin, Austin, TX, April
1988.
[165] J. M. ORTEGA, Introduction to Parallel and Vector Solution of Linear Systems, Plenum
Press, New York and London, 1988.
[166] C. PAIGE, B. PARLETT, AND H. VAN DER VORST, Approximate solutions and eigenvalue
bounds from Krylov subspaces, Numer. Lin. Alg. Appls., to appear.
[167] C. PAIGE AND M. SAUNDERS, Solution of sparse indefinite systems of linear equations,
SIAM J. Numer. Anal., 12 (1975), pp. 617–629.
[168] C. C. PAIGE AND M. A. SAUNDERS, LSQR: An algorithm for sparse linear equations and
sparse least squares, ACM Trans. Math. Soft., 8 (1982), pp. 43–71.
[169] G. PAOLINI AND G. RADICATI DI BROZOLO, Data structures to vectorize CG algorithms
for general sparsity patterns, BIT, 29 (1989), pp. 703–718.
[170] B. PARLETT, The symmetric eigenvalue problem, Prentice-Hall, London, 1980.
[171] B. N. PARLETT, D. R. TAYLOR, AND Z. A. LIU, A look-ahead Lanczos algorithm for
unsymmetric matrices, Mathematics of Computation, 44 (1985), pp. 105–124.
[172] D. PEACEMAN AND J. H.H. RACHFORD, The numerical solution of parabolic and elliptic
differential equations, J. Soc. Indust. Appl. Math., 3 (1955), pp. 28–41.
[173] C. POMMERELL, Solution of Large Unsymmetric Systems of Linear Equations, vol. 17 of
Series in Micro-electronics, volume 17, Hartung-Gorre Verlag, Konstanz, 1992.
[174] , Solution of large unsymmetric systems of linear equations, PhD thesis, Swiss Federal
Institute of Technology, Z¨urich, Switzerland, 1992.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-112-320.jpg)
![BIBLIOGRAPHY 103
[175] E. POOLE AND J. ORTEGA, Multicolor ICCG methods for vector computers, Tech. Rep.
RM 86-06, Department of Applied Mathematics, University of Virginia, Charlottesville,
VA, 1986.
[176] A. QUARTERONI, ed., Domain Decomposition Methods, Proceedings of the Sixth Interna-
tional Symposium on Domain Decomposition Methods, Como, Italy,, Providence, RI, 1993,
AMS. to appear.
[177] G. RADICATI DI BROZOLO AND Y. ROBERT, Vector and parallel CG-like algorithms for
sparse non-symmetric systems, Tech. Rep. 681-M, IMAG/TIM3, Grenoble, France, 1987.
[178] J. REID, On the method of conjugate gradients for the solution of large sparse systems of
linear equations, in Large Sparse Sets of Linear Equations, J. Reid, ed., Academic Press,
London, 1971, pp. 231–254.
[179] G. RODRIGUE AND D. WOLITZER, Preconditioning by incomplete block cyclic reduction,
Mathematics of Computation, 42 (1984), pp. 549–565.
[180] Y. SAAD, The Lanczos biorthogonalization algorithm and other oblique projection methods
for solving large unsymmetric systems, SIAM J. Numer. Anal., 19 (1982), pp. 485–506.
[181] , Practical use of some Krylov subspace methods for solving indefinite and nonsym-
metric linear systems, SIAM J. Sci. Statist. Comput., 5 (1984), pp. 203–228.
[182] , Practical use of polynomial preconditionings for the conjugate gradient method,
SIAM J. Sci. Statist. Comput., 6 (1985), pp. 865–881.
[183] , Preconditioning techniques for indefinite and nonsymmetric linear systems, J. Com-
put. Appl. Math., 24 (1988), pp. 89–105.
[184] , Krylov subspace methods on supercomputers, SIAM J. Sci. Statist. Comput., 10
(1989), pp. 1200–1232.
[185] , SPARSKIT: A basic tool kit for sparse matrix computation, Tech. Rep. CSRD TR
1029, CSRD, University of Illinois, Urbana, IL, 1990.
[186] , A flexible inner-outer preconditioned GMRES algorithm, SIAM J. Sci. Comput., 14
(1993), pp. 461–469.
[187] Y. SAAD AND M. SCHULTZ, Conjugate gradient-like algorithms for solving nonsymmetric
linear systems, Mathematics of Computation, 44 (1985), pp. 417–424.
[188] , GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear
systems, SIAM J. Sci. Statist. Comput., 7 (1986), pp. 856–869.
[189] G. L. G. SLEIJPEN AND D. R. FOKKEMA, Bi-CGSTAB( ) for linear equations involving
unsymmetric matrices with complex spectrum, Tech. Rep. 772, University of Utrecht, De-
partment of Mathematics, Utrecht, The Netherlands, 1993.
[190] B. F. SMITH, Domain decomposition algorithms for partial differential equations of linear
elasticity, Tech. Rep. 517, Department of Computer Science, Courant Institute, 1990.
[191] P. SONNEVELD, CGS, a fast Lanczos-type solver for nonsymmetric linear systems, SIAM J.
Sci. Statist. Comput., 10 (1989), pp. 36–52.
[192] R. SOUTHWELL, Relaxation Methods in Theoretical Physics, Clarendon Press, Oxford,
1946.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-113-320.jpg)
![104 BIBLIOGRAPHY
[193] H. STONE, Iterative solution of implicit approximations of multidimensional partial differ-
ential equations, SIAM J. Numer. Anal., 5 (1968), pp. 530–558.
[194] P. SWARZTRAUBER, The methods of cyclic reduction, Fourier analysis and the FACR algo-
rithm for the discrete solution of Poisson’s equation on a rectangle, SIAM Rev., 19 (1977),
pp. 490–501.
[195] C. TONG, A comparative study of preconditioned Lanczos methods for nonsymmetric linear
systems, Tech. Rep. SAND91-8240, Sandia Nat. Lab., Livermore, CA, 1992.
[196] A. VAN DER SLUIS, Condition numbers and equilibration of matrices, Numer. Math., 14
(1969), pp. 14–23.
[197] A. VAN DER SLUIS AND H. VAN DER VORST, The rate of convergence of conjugate gradi-
ents, Numer. Math., 48 (1986), pp. 543–560.
[198] H. VAN DER VORST, Iterative solution methods for certain sparse linear systems with a
non-symmetric matrix arising from PDE-problems, J. Comput. Phys., 44 (1981), pp. 1–19.
[199] , A vectorizable variant of some ICCG methods, SIAM J. Sci. Statist. Comput., 3
(1982), pp. 350–356.
[200] , Large tridiagonal and block tridiagonal linear systems on vector and parallel com-
puters, Parallel Comput., 5 (1987), pp. 45–54.
[201] , (M)ICCG for 2D problems on vector computers, in Supercomputing, A.Lichnewsky
and C.Saguez, eds., North-Holland, 1988. Also as Report No.A-17, Data Processing Center,
Kyoto University, Kyoto, Japan, December 17, 1986.
[202] , High performance preconditioning, SIAM J. Sci. Statist. Comput., 10 (1989),
pp. 1174–1185.
[203] , ICCG and related methods for 3D problems on vector computers, Computer Physics
Communications, (1989), pp. 223–235. Also as Report No.A-18, Data Processing Center,
Kyoto University, Kyoto, Japan, May 30, 1987.
[204] , The convergence behavior of preconditioned CG and CG-S in the presence of round-
ing errors, in Preconditioned Conjugate Gradient Methods, O. Axelsson and L. Y. Kolotilina,
eds., vol. 1457 of Lecture Notes in Mathematics, Berlin, New York, 1990, Springer-Verlag.
[205] , Bi-CGSTAB: A fast and smoothly converging variant of Bi-CG for the solution of
nonsymmetric linear systems, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 631–644.
[206] H. VAN DER VORST AND J. MELISSEN, A Petrov-Galerkin type method for solving Ax = b
where A is symmetric complex, IEEE Trans. Magnetics, 26 (1990), pp. 706–708.
[207] H. VAN DER VORST AND C. VUIK, GMRESR: A family of nested GMRES methods, Tech.
Rep. 91-80, Delft University of Technology, Faculty of Tech. Math., Delft, The Netherlands,
1991.
[208] J. VAN ROSENDALE, Minimizing inner product data dependencies in conjugate gradient
iteration, Tech. Rep. 172178, ICASE, NASA Langley Research Center, 1983.
[209] R. VARGA, Matrix Iterative Analysis, Prentice-Hall Inc., Englewood Cliffs, NJ, 1962.
[210] P. VASSILEVSKI, Preconditioning nonsymmetric and indefinite finite element matrices, J.
Numer. Alg. Appl., 1 (1992), pp. 59–76.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-114-320.jpg)
![BIBLIOGRAPHY 105
[211] V. VOEVODIN, The problem of non-self-adjoint generalization of the conjugate gradient
method is closed, U.S.S.R. Comput. Maths. and Math. Phys., 23 (1983), pp. 143–144.
[212] H. F. WALKER, Implementation of the GMRES method using Householder transformations,
SIAM J. Sci. Statist. Comput., 9 (1988), pp. 152–163.
[213] P. WESSELING, An Introduction to Multigrid Methods, Wiley, Chichester, 1991.
[214] O. WIDLUND, A Lanczos method for a class of non-symmetric systems of linear equations,
SIAM J. Numer. Anal., 15 (1978), pp. 801–812.
[215] D. YOUNG, Iterative solution of large linear systems, Academic Press, New York, 1971.
[216] H. YSERENTANT, On the multilevel splitting of finite element spaces, Numer. Math., 49
(1986), pp. 379–412.](https://image.slidesharecdn.com/barrettemplates-140514004551-phpapp02/85/Barret-templates-115-320.jpg)
Barret templates
- 1. Templates for the Solution of Linear Systems: Building Blocks for Iterative Methods1 Richard Barrett2 , Michael Berry3 , Tony F. Chan4 , James Demmel5 , June M. Donato6 , Jack Dongarra3,2 , Victor Eijkhout7 , Roldan Pozo8 , Charles Romine9 , and Henk Van der Vorst10 This document is the electronic version of the 2nd edition of the Templates book, which is available for purchase from the Society for Industrial and Applied Mathematics (http://www.siam.org/books). 1 This work was supported in part by DARPA and ARO under contract number DAAL03-91-C-0047, the National Science Foundation Science and Technology Center Cooperative Agreement No. CCR-8809615, the Applied Mathematical Sciences subprogram of the Office of Energy Research, U.S. Department of Energy, under Contract DE-AC05-84OR21400, and the Stichting Nationale Computer Faciliteit (NCF) by Grant CRG 92.03. 2 Computer Science and Mathematics Division, Oak Ridge National Laboratory, Oak Ridge, TN 37830- 6173. 3 Department of Computer Science, University of Tennessee, Knoxville, TN 37996. 4 Applied Mathematics Department, University of California, Los Angeles, CA 90024-1555. 5 Computer Science Division and Mathematics Department, University of California, Berkeley, CA 94720. 6 Science Applications International Corporation, Oak Ridge, TN 37831 7 Texas Advanced Computing Center, The University of Texas at Austin, Austin, TX 78758 8 National Institute of Standards and Technology, Gaithersburg, MD 9 Office of Science and Technology Policy, Executive Office of the President 10 Department of Mathematics, Utrecht University, Utrecht, the Netherlands.
- 2. ii
- 3. How to Use This Book We have divided this book into five main chapters. Chapter 1 gives the motivation for this book and the use of templates. Chapter 2 describes stationary and nonstationary iterative methods. In this chapter we present both historical development and state-of-the-art methods for solving some of the most challenging computational problems facing researchers. Chapter 3 focuses on preconditioners. Many iterative methods depend in part on precondition- ers to improve performance and ensure fast convergence. Chapter 4 provides a glimpse of issues related to the use of iterative methods. This chapter, like the preceding, is especially recommended for the experienced user who wishes to have further guidelines for tailoring a specific code to a particular machine. It includes information on complex systems, stopping criteria, data storage formats, and parallelism. Chapter 5 includes overviews of related topics such as the close connection between the Lanc- zos algorithm and the Conjugate Gradient algorithm, block iterative methods, red/black orderings, domain decomposition methods, multigrid-like methods, and row-projection schemes. The Appendices contain information on how the templates and BLAS software can be obtained. A glossary of important terms used in the book is also provided. The field of iterative methods for solving systems of linear equations is in constant flux, with new methods and approaches continually being created, modified, tuned, and some eventually discarded. We expect the material in this book to undergo changes from time to time as some of these new approaches mature and become the state-of-the-art. Therefore, we plan to update the material included in this book periodically for future editions. We welcome your comments and criticisms of this work to help us in that updating process. Please send your comments and questions by email to [email protected]. iii
- 4. List of Symbols A, . . . , Z matrices a, . . . , z vectors α, β, . . . , ω scalars AT matrix transpose AH conjugate transpose (Hermitian) of A A−1 matrix inverse A−T the inverse of AT ai,j matrix element a.,j jth matrix column Ai,j matrix subblock ai vector element ux, uxx first, second derivative with respect to x (x, y), xT y vector dot product (inner product) x (i) j jth component of vector x in the ith iteration diag(A) diagonal of matrix A diag(α, β . . .) diagonal matrix constructed from scalars α, β . . . span(a, b . . .) spanning space of vectors a, b . . . R set of real numbers Rn real n-space ||x||2 2-norm ||x||p p-norm ||x||A the “A-norm”, defined as (Ax, x)1/2 λmax(A), λmin(A) eigenvalues of A with maximum (resp. minimum) modulus σmax(A), σmin(A) largest and smallest singular values of A κ2(A) spectral condition number of matrix A L linear operator α complex conjugate of the scalar α max{S} maximum value in set S min{S} minimum value in set S summation O(·) “big-oh” asymptotic bound Conventions Used in this Book D diagonal matrix L lower triangular matrix U upper triangular matrix Q orthogonal matrix M preconditioner I, In×n n × n identity matrix ˆx typically, the exact solution to Ax = b h discretization mesh width iv
- 5. Acknowledgements The authors gratefully acknowledge the valuable assistance of many people who commented on preliminary drafts of this book. In particular, we thank Loyce Adams, Bill Coughran, Matthew Fishler, Peter Forsyth, Roland Freund, Gene Golub, Eric Grosse, Mark Jones, David Kincaid, Steve Lee, Tarek Mathew, No¨el Nachtigal, Jim Ortega, and David Young for their insightful com- ments. We also thank Geoffrey Fox for initial discussions on the concept of templates, and Karin Remington for designing the front cover. This work was supported in part by DARPA and ARO under contract number DAAL03-91- C-0047, the National Science Foundation Science and Technology Center Cooperative Agreement No. CCR-8809615, the Applied Mathematical Sciences subprogram of the Office of Energy Re- search, U.S. Department of Energy, under Contract DE-AC05-84OR21400, and the Stichting Na- tionale Computer Faciliteit (NCF) by Grant CRG 92.03. v
- 6. vi
- 7. Contents List of Symbols iv List of Figures ix 1 Introduction 1 1.1 Why Use Templates? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.2 What Methods Are Covered? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 Iterative Methods 5 2.1 Overview of the Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.2 Stationary Iterative Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.1 The Jacobi Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.2 The Gauss-Seidel Method . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2.3 The Successive Overrelaxation Method . . . . . . . . . . . . . . . . . . . 9 2.2.4 The Symmetric Successive Overrelaxation Method . . . . . . . . . . . . . 11 2.2.5 Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.3 Nonstationary Iterative Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.3.1 Conjugate Gradient Method (CG) . . . . . . . . . . . . . . . . . . . . . . 12 2.3.2 MINRES and SYMMLQ . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3.3 CG on the Normal Equations, CGNE and CGNR . . . . . . . . . . . . . . 16 2.3.4 Generalized Minimal Residual (GMRES) . . . . . . . . . . . . . . . . . . 17 2.3.5 BiConjugate Gradient (BiCG) . . . . . . . . . . . . . . . . . . . . . . . . 19 2.3.6 Quasi-Minimal Residual (QMR) . . . . . . . . . . . . . . . . . . . . . . . 20 2.3.7 Conjugate Gradient Squared Method (CGS) . . . . . . . . . . . . . . . . . 21 2.3.8 BiConjugate Gradient Stabilized (Bi-CGSTAB) . . . . . . . . . . . . . . . 24 2.3.9 Chebyshev Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.4 Computational Aspects of the Methods . . . . . . . . . . . . . . . . . . . . . . . . 26 2.5 A short history of Krylov methods . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.6 Survey of recent Krylov methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3 Preconditioners 35 3.1 The why and how . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 3.1.1 Cost trade-off . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 3.1.2 Left and right preconditioning . . . . . . . . . . . . . . . . . . . . . . . . 36 3.2 Jacobi Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.2.1 Block Jacobi Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.2.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.3 SSOR preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.4 Incomplete Factorization Preconditioners . . . . . . . . . . . . . . . . . . . . . . 38 3.4.1 Creating an incomplete factorization . . . . . . . . . . . . . . . . . . . . . 38 vii
- 8. viii CONTENTS 3.4.2 Point incomplete factorizations . . . . . . . . . . . . . . . . . . . . . . . . 39 3.4.3 Block factorization methods . . . . . . . . . . . . . . . . . . . . . . . . . 44 3.4.4 Blocking over systems of partial differential equations . . . . . . . . . . . 46 3.4.5 Incomplete LQ factorizations . . . . . . . . . . . . . . . . . . . . . . . . 46 3.5 Polynomial preconditioners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.6 Other preconditioners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.6.1 Preconditioning by the symmetric part . . . . . . . . . . . . . . . . . . . . 47 3.6.2 The use of fast solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.6.3 Alternating Direction Implicit methods . . . . . . . . . . . . . . . . . . . 48 4 Related Issues 51 4.1 Complex Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 4.2 Stopping Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 4.2.1 More Details about Stopping Criteria . . . . . . . . . . . . . . . . . . . . 52 4.2.2 When r(i) or r(i) is not readily available . . . . . . . . . . . . . . . . . 55 4.2.3 Estimating A−1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.2.4 Stopping when progress is no longer being made . . . . . . . . . . . . . . 56 4.2.5 Accounting for floating point errors . . . . . . . . . . . . . . . . . . . . . 56 4.3 Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 4.3.1 Survey of Sparse Matrix Storage Formats . . . . . . . . . . . . . . . . . . 57 4.3.2 Matrix vector products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 4.3.3 Sparse Incomplete Factorizations . . . . . . . . . . . . . . . . . . . . . . 63 4.4 Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 4.4.1 Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 4.4.2 Vector updates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.4.3 Matrix-vector products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.4.4 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.4.5 Wavefronts in the Gauss-Seidel and Conjugate Gradient methods . . . . . 70 4.4.6 Blocked operations in the GMRES method . . . . . . . . . . . . . . . . . 71 5 Remaining topics 73 5.1 The Lanczos Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 5.2 Block and s-step Iterative Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 74 5.3 Reduced System Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 5.4 Domain Decomposition Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 5.4.1 Overlapping Subdomain Methods . . . . . . . . . . . . . . . . . . . . . . 76 5.4.2 Non-overlapping Subdomain Methods . . . . . . . . . . . . . . . . . . . . 77 5.4.3 Further Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.5 Multigrid Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 5.6 Row Projection Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 A Obtaining the Software 83 B Overview of the BLAS 85 C Glossary 87 C.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 D Flowchart of iterative methods 107
- 9. List of Figures 2.1 The Jacobi Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2 The Gauss-Seidel Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.3 The SOR Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4 The SSOR Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.5 The Preconditioned Conjugate Gradient Method . . . . . . . . . . . . . . . . . . . 13 2.6 The Preconditioned GMRES(m) Method . . . . . . . . . . . . . . . . . . . . . . 18 2.7 The Preconditioned BiConjugate Gradient Method . . . . . . . . . . . . . . . . . 20 2.8 The Preconditioned Quasi Minimal Residual Method without Look-ahead . . . . . 22 2.9 The Preconditioned Conjugate Gradient Squared Method . . . . . . . . . . . . . . 23 2.10 The Preconditioned BiConjugate Gradient Stabilized Method . . . . . . . . . . . . 24 2.11 The Preconditioned Chebyshev Method . . . . . . . . . . . . . . . . . . . . . . . 26 3.1 Preconditioner solve of a system Mx = y, with M = LU . . . . . . . . . . . . . 39 3.2 Preconditioner solve of a system Mx = y, with M = (D + L)D−1 (D + U) = (D + L)(I + D−1 U). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.3 Construction of a D-ILU incomplete factorization preconditioner, storing the in- verses of the pivots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.4 Wavefront solution of (D+L)x = u from a central difference problem on a domain of n × n points. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.5 Preconditioning step algorithm for a Neumann expansion M(p) ≈ M−1 of an incomplete factorization M = (I + L)D(I + U). . . . . . . . . . . . . . . . . . . 43 3.6 Block version of a D-ILU factorization . . . . . . . . . . . . . . . . . . . . . . . 44 3.7 Algorithm for approximating the inverse of a banded matrix . . . . . . . . . . . . 45 3.8 Incomplete block factorization of a block tridiagonal matrix . . . . . . . . . . . . 45 4.1 Profile of a nonsymmetric skyline or variable-band matrix. . . . . . . . . . . . . . 61 4.2 A rearrangement of Conjugate Gradient for parallelism . . . . . . . . . . . . . . . 68 ix
- 10. x LIST OF FIGURES
- 11. Chapter 1 Introduction Which of the following statements is true? • Users want “black box” software that they can use with complete confidence for general problem classes without having to understand the fine algorithmic details. • Users want to be able to tune data structures for a particular application, even if the software is not as reliable as that provided for general methods. It turns out both are true, for different groups of users. Traditionally, users have asked for and been provided with black box software in the form of mathematical libraries such as LAPACK, LINPACK, NAG, and IMSL. More recently, the high- performance community has discovered that they must write custom software for their problem. Their reasons include inadequate functionality of existing software libraries, data structures that are not natural or convenient for a particular problem, and overly general software that sacrifices too much performance when applied to a special case of interest. Can we meet the needs of both groups of users? We believe we can. Accordingly, in this book, we introduce the use of templates. A template is a description of a general algorithm rather than the executable object code or the source code more commonly found in a conventional software library. Nevertheless, although templates are general descriptions of key algorithms, they offer whatever degree of customization the user may desire. For example, they can be configured for the specific data structure of a problem or for the specific computing system on which the problem is to run. We focus on the use of iterative methods for solving large sparse systems of linear equations. Many methods exist for solving such problems. The trick is to find the most effective method for the problem at hand. Unfortunately, a method that works well for one problem type may not work as well for another. Indeed, it may not work at all. Thus, besides providing templates, we suggest how to choose and implement an effective method, and how to specialize a method to specific matrix types. We restrict ourselves to iter- ative methods, which work by repeatedly improving an approximate solution until it is accurate enough. These methods access the coefficient matrix A of the linear system only via the matrix- vector product y = A · x (and perhaps z = AT · x). Thus the user need only supply a subroutine for computing y (and perhaps z) given x, which permits full exploitation of the sparsity or other special structure of A. We believe that after reading this book, applications developers will be able to use templates to get their program running on a parallel machine quickly. Nonspecialists will know how to choose and implement an approach to solve a particular problem. Specialists will be able to assemble and modify their codes—without having to make the huge investment that has, up to now, been required to tune large-scale applications for each particular machine. Finally, we hope that all users will gain a better understanding of the algorithms employed. While education has not been one of 1
- 12. 2 CHAPTER 1. INTRODUCTION the traditional goals of mathematical software, we believe that our approach will go a long way in providing such a valuable service. 1.1 Why Use Templates? Templates offer three significant advantages. First, templates are general and reusable. Thus, they can simplify ports to diverse machines. This feature is important given the diversity of parallel architectures. Second, templates exploit the expertise of two distinct groups. The expert numerical analyst creates a template reflecting in-depth knowledge of a specific numerical technique. The compu- tational scientist then provides “value-added” capability to the general template description, cus- tomizing it for specific contexts or applications needs. And third, templates are not language specific. Rather, they are displayed in an Algol-like structure, which is readily translatable into the target language such as FORTRAN (with the use of the Basic Linear Algebra Subprograms, or BLAS, whenever possible) and C. By using these familiar styles, we believe that the users will have trust in the algorithms. We also hope that users will gain a better understanding of numerical techniques and parallel programming. For each template, we provide some or all of the following: • a mathematical description of the flow of the iteration; • discussion of convergence and stopping criteria; • suggestions for applying a method to special matrix types (e.g., banded systems); • advice for tuning (for example, which preconditioners are applicable and which are not); • tips on parallel implementations; and • hints as to when to use a method, and why. For each of the templates, the following can be obtained via electronic mail. • a MATLAB implementation based on dense matrices; • a FORTRAN-77 program with calls to BLAS1 . See Appendix A for details. 1.2 What Methods Are Covered? Many iterative methods have been developed and it is impossible to cover them all. We chose the methods below either because they illustrate the historical development of iterative methods, or because they represent the current state-of-the-art for solving large sparse linear systems. The methods we discuss are: 1. Jacobi 2. Gauss-Seidel 3. Successive Over-Relaxation (SOR) 4. Symmetric Successive Over-Relaxation (SSOR) 1For a discussion of BLAS as building blocks, see [68, 69, 70, 143] and LAPACK routines [3]. Also, see Appendix B.
- 13. 1.2. WHAT METHODS ARE COVERED? 3 5. Conjugate Gradient (CG) 6. Minimal Residual (MINRES) and Symmetric LQ (SYMMLQ) 7. Conjugate Gradients on the Normal Equations (CGNE and CGNR) 8. Generalized Minimal Residual (GMRES) 9. Biconjugate Gradient (BiCG) 10. Quasi-Minimal Residual (QMR) 11. Conjugate Gradient Squared (CGS) 12. Biconjugate Gradient Stabilized (Bi-CGSTAB) 13. Chebyshev Iteration For each method we present a general description, including a discussion of the history of the method and numerous references to the literature. We also give the mathematical conditions for selecting a given method. We do not intend to write a “cookbook”, and have deliberately avoided the words “numerical recipes”, because these phrases imply that our algorithms can be used blindly without knowledge of the system of equations. The state of the art in iterative methods does not permit this: some knowledge about the linear system is needed to guarantee convergence of these algorithms, and generally the more that is known the more the algorithm can be tuned. Thus, we have chosen to present an algorithmic outline, with guidelines for choosing a method and implementing it on particular kinds of high-performance machines. We also discuss the use of preconditioners and relevant data storage issues.
- 14. 4 CHAPTER 1. INTRODUCTION
- 15. Chapter 2 Iterative Methods The term “iterative method” refers to a wide range of techniques that use successive approximations to obtain more accurate solutions to a linear system at each step. In this book we will cover two types of iterative methods. Stationary methods are older, simpler to understand and implement, but usually not as effective. Nonstationary methods are a relatively recent development; their analysis is usually harder to understand, but they can be highly effective. The nonstationary methods we present are based on the idea of sequences of orthogonal vectors. (An exception is the Chebyshev iteration method, which is based on orthogonal polynomials.) The rate at which an iterative method converges depends greatly on the spectrum of the co- efficient matrix. Hence, iterative methods usually involve a second matrix that transforms the coefficient matrix into one with a more favorable spectrum. The transformation matrix is called a preconditioner. A good preconditioner improves the convergence of the iterative method, suffi- ciently to overcome the extra cost of constructing and applying the preconditioner. Indeed, without a preconditioner the iterative method may even fail to converge. 2.1 Overview of the Methods Below are short descriptions of each of the methods to be discussed, along with brief notes on the classification of the methods in terms of the class of matrices for which they are most appropriate. In later sections of this chapter more detailed descriptions of these methods are given. • Stationary Methods – Jacobi. The Jacobi method is based on solving for every variable locally with respect to the other variables; one iteration of the method corresponds to solving for every variable once. The resulting method is easy to understand and implement, but convergence is slow. – Gauss-Seidel. The Gauss-Seidel method is like the Jacobi method, except that it uses updated values as soon as they are available. In general, if the Jacobi method converges, the Gauss-Seidel method will converge faster than the Jacobi method, though still relatively slowly. – SOR. Successive Overrelaxation (SOR) can be derived from the Gauss-Seidel method by introducing an extrapolation parameter ω. For the optimal choice of ω, SOR may con- verge faster than Gauss-Seidel by an order of magnitude. 5
- 16. 6 CHAPTER 2. ITERATIVE METHODS – SSOR. Symmetric Successive Overrelaxation (SSOR) has no advantage over SOR as a stand- alone iterative method; however, it is useful as a preconditioner for nonstationary meth- ods. • Nonstationary Methods – Conjugate Gradient (CG). The conjugate gradient method derives its name from the fact that it generates a se- quence of conjugate (or orthogonal) vectors. These vectors are the residuals of the iter- ates. They are also the gradients of a quadratic functional, the minimization of which is equivalent to solving the linear system. CG is an extremely effective method when the coefficient matrix is symmetric positive definite, since storage for only a limited number of vectors is required. – Minimum Residual (MINRES) and Symmetric LQ (SYMMLQ). These methods are computational alternatives for CG for coefficient matrices that are symmetric but possibly indefinite. SYMMLQ will generate the same solution iterates as CG if the coefficient matrix is symmetric positive definite. – Conjugate Gradient on the Normal Equations: CGNE and CGNR. These methods are based on the application of the CG method to one of two forms of the normal equations for Ax = b. CGNE solves the system (AAT )y = b for y and then computes the solution x = AT y. CGNR solves (AT A)x = ˜b for the solution vector x where ˜b = AT b. When the coefficient matrix A is nonsymmetric and nonsingular, the normal equations matrices AAT and AT A will be symmetric and positive definite, and hence CG can be applied. The convergence may be slow, since the spectrum of the normal equations matrices will be less favorable than the spectrum of A. – Generalized Minimal Residual (GMRES). The Generalized Minimal Residual method computes a sequence of orthogonal vectors (like MINRES), and combines these through a least-squares solve and update. How- ever, unlike MINRES (and CG) it requires storing the whole sequence, so that a large amount of storage is needed. For this reason, restarted versions of this method are used. In restarted versions, computation and storage costs are limited by specifying a fixed number of vectors to be generated. This method is useful for general nonsymmetric matrices. – BiConjugate Gradient (BiCG). The Biconjugate Gradient method generates two CG-like sequences of vectors, one based on a system with the original coefficient matrix A, and one on AT . Instead of orthogonalizing each sequence, they are made mutually orthogonal, or “bi-orthogonal”. This method, like CG, uses limited storage. It is useful when the matrix is nonsymmet- ric and nonsingular; however, convergence may be irregular, and there is a possibility that the method will break down. BiCG requires a multiplication with the coefficient matrix and with its transpose at each iteration. – Quasi-Minimal Residual (QMR). The Quasi-Minimal Residual method applies a least-squares solve and update to the BiCG residuals, thereby smoothing out the irregular convergence behavior of BiCG. Also, QMR largely avoids the breakdown that can occur in BiCG. On the other hand, it does not effect a true minimization of either the error or the residual, and while it converges smoothly, it does not essentially improve on the BiCG. – Conjugate Gradient Squared (CGS).
- 17. 2.2. STATIONARY ITERATIVE METHODS 7 The Conjugate Gradient Squared method is a variant of BiCG that applies the updating operations for the A-sequence and the AT -sequences both to the same vectors. Ideally, this would double the convergence rate, but in practice convergence may be much more irregular than for BiCG. A practical advantage is that the method does not need the multiplications with the transpose of the coefficient matrix. – Biconjugate Gradient Stabilized (Bi-CGSTAB). The Biconjugate Gradient Stabilized method is a variant of BiCG, like CGS, but using different updates for the AT -sequence in order to obtain smoother convergence than CGS. – Chebyshev Iteration. The Chebyshev Iteration recursively determines polynomials with coefficients chosen to minimize the norm of the residual in a min-max sense. The coefficient matrix must be positive definite and knowledge of the extremal eigenvalues is required. This method has the advantage of requiring no inner products. 2.2 Stationary Iterative Methods Iterative methods that can be expressed in the simple form x(k) = Bx(k−1) + c (2.1) (where neither B nor c depend upon the iteration count k) are called stationary iterative methods. In this section, we present the four main stationary iterative methods: the Jacobi method, the Gauss-Seidel method, the Successive Overrelaxation (SOR) method and the Symmetric Successive Overrelaxation (SSOR) method. In each case, we summarize their convergence behavior and their effectiveness, and discuss how and when they should be used. Finally, in §2.2.5, we give some historical background and further notes and references. 2.2.1 The Jacobi Method The Jacobi method is easily derived by examining each of the n equations in the linear system Ax = b in isolation. If in the ith equation n j=1 ai,jxj = bi, we solve for the value of xi while assuming the other entries of x remain fixed, we obtain xi = (bi − j=i ai,jxj)/ai,i. (2.2) This suggests an iterative method defined by x (k) i = (bi − j=i ai,jx (k−1) j )/ai,i, (2.3) which is the Jacobi method. Note that the order in which the equations are examined is irrelevant, since the Jacobi method treats them independently. For this reason, the Jacobi method is also known as the method of simultaneous displacements, since the updates could in principle be done simultaneously. In matrix terms, the definition of the Jacobi method in (2.3) can be expressed as x(k) = D−1 (L + U)x(k−1) + D−1 b, (2.4)
- 18. 8 CHAPTER 2. ITERATIVE METHODS Choose an initial guess x(0) to the solution x. for k = 1, 2, . . . for i = 1, 2, . . . , n ¯xi = 0 for j = 1, 2, . . . , i − 1, i + 1, . . . , n ¯xi = ¯xi + ai,jx (k−1) j end ¯xi = (bi − ¯xi)/ai,i end x(k) = ¯x check convergence; continue if necessary end Figure 2.1: The Jacobi Method where the matrices D, −L and −U represent the diagonal, the strictly lower-triangular, and the strictly upper-triangular parts of A, respectively. The pseudocode for the Jacobi method is given in Figure 2.1. Note that an auxiliary storage vector, ¯x is used in the algorithm. It is not possible to update the vector x in place, since values from x(k−1) are needed throughout the computation of x(k) . Convergence of the Jacobi method Iterative methods are often used for solving discretized partial differential equations. In that context a rigorous analysis of the convergence of simple methods such as the Jacobi method can be given. As an example, consider the boundary value problem Lu = −uxx = f on (0, 1), u(0) = u0, u(1) = u1, discretized by Lu(xi) = 2u(xi) − u(xi−1) − u(xi+1) = f(xi)/N2 for xi = i/N, i = 1 . . . N − 1. The eigenfunctions of the L and L operator are the same: for n = 1 . . . N − 1 the function un(x) = sin nπx is an eigenfunction corresponding to λ = 4 sin2 nπ/(2N). The eigenvalues of the Jacobi iteration matrix B are then λ(B) = 1 − 1/2λ(L) = 1 − 2 sin2 nπ/(2N). From this it is easy to see that the high frequency modes (i.e., eigenfunction un with n large) are damped quickly, whereas the damping factor for modes with n small is close to 1. The spectral radius of the Jacobi iteration matrix is ≈ 1 − 10/N2 , and it is attained for the eigenfunction u(x) = sin πx. The type of analysis applied to this example can be generalized to higher dimensions and other stationary iterative methods. For both the Jacobi and Gauss-Seidel method (below) the spectral radius is found to be 1 − O(h2 ) where h is the discretization mesh width, i.e., h = N−d where N is the number of variables and d is the number of space dimensions. 2.2.2 The Gauss-Seidel Method Consider again the linear equations in (2.2). If we proceed as with the Jacobi method, but now assume that the equations are examined one at a time in sequence, and that previously computed
- 19. 2.2. STATIONARY ITERATIVE METHODS 9 Choose an initial guess x(0) to the solution x. for k = 1, 2, . . . for i = 1, 2, . . . , n σ = 0 for j = 1, 2, . . . , i − 1 σ = σ + ai,jx (k) j end for j = i + 1, . . . , n σ = σ + ai,jx (k−1) j end x (k) i = (bi − σ)/ai,i end check convergence; continue if necessary end Figure 2.2: The Gauss-Seidel Method results are used as soon as they are available, we obtain the Gauss-Seidel method: x (k) i = (bi − j<i ai,jx (k) j − j>i ai,jx (k−1) j )/ai,i. (2.5) Two important facts about the Gauss-Seidel method should be noted. First, the computations in (2.5) appear to be serial. Since each component of the new iterate depends upon all previ- ously computed components, the updates cannot be done simultaneously as in the Jacobi method. Second, the new iterate x(k) depends upon the order in which the equations are examined. The Gauss-Seidel method is sometimes called the method of successive displacements to indicate the dependence of the iterates on the ordering. If this ordering is changed, the components of the new iterate (and not just their order) will also change. These two points are important because if A is sparse, the dependency of each component of the new iterate on previous components is not absolute. The presence of zeros in the matrix may remove the influence of some of the previous components. Using a judicious ordering of the equations, it may be possible to reduce such dependence, thus restoring the ability to make updates to groups of components in parallel. However, reordering the equations can affect the rate at which the Gauss-Seidel method converges. A poor choice of ordering can degrade the rate of convergence; a good choice can enhance the rate of convergence. For a practical discussion of this tradeoff (parallelism versus convergence rate) and some standard reorderings, the reader is referred to Chapter 3 and §4.4. In matrix terms, the definition of the Gauss-Seidel method in (2.5) can be expressed as x(k) = (D − L)−1 (Ux(k−1) + b). (2.6) As before, D, −L and −U represent the diagonal, lower-triangular, and upper-triangular parts of A, respectively. The pseudocode for the Gauss-Seidel algorithm is given in Figure 2.2. 2.2.3 The Successive Overrelaxation Method The Successive Overrelaxation Method, or SOR, is devised by applying extrapolation to the Gauss- Seidel method. This extrapolation takes the form of a weighted average between the previous iterate
- 20. 10 CHAPTER 2. ITERATIVE METHODS Choose an initial guess x(0) to the solution x. for k = 1, 2, . . . for i = 1, 2, . . . , n σ = 0 for j = 1, 2, . . . , i − 1 σ = σ + ai,jx (k) j end for j = i + 1, . . . , n σ = σ + ai,jx (k−1) j end σ = (bi − σ)/ai,i x (k) i = x (k−1) i + ω(σ − x (k−1) i ) end check convergence; continue if necessary end Figure 2.3: The SOR Method and the computed Gauss-Seidel iterate successively for each component: x (k) i = ω¯x (k) i + (1 − ω)x (k−1) i (where ¯x denotes a Gauss-Seidel iterate, and ω is the extrapolation factor). The idea is to choose a value for ω that will accelerate the rate of convergence of the iterates to the solution. In matrix terms, the SOR algorithm can be written as follows: x(k) = (D − ωL)−1 (ωU + (1 − ω)D)x(k−1) + ω(D − ωL)−1 b. (2.7) The pseudocode for the SOR algorithm is given in Figure 2.3. Choosing the Value of ω If ω = 1, the SOR method simplifies to the Gauss-Seidel method. A theorem due to Kahan [129] shows that SOR fails to converge if ω is outside the interval (0, 2). Though technically the term underrelaxation should be used when 0 < ω < 1, for convenience the term overrelaxation is now used for any value of ω ∈ (0, 2). In general, it is not possible to compute in advance the value of ω that is optimal with respect to the rate of convergence of SOR. Even when it is possible to compute the optimal value for ω, the expense of such computation is usually prohibitive. Frequently, some heuristic estimate is used, such as ω = 2 − O(h) where h is the mesh spacing of the discretization of the underlying physical domain. If the coefficient matrix A is symmetric and positive definite, the SOR iteration is guaranteed to converge for any value of ω between 0 and 2, though the choice of ω can significantly affect the rate at which the SOR iteration converges. Sophisticated implementations of the SOR algorithm (such as that found in ITPACK [139]) employ adaptive parameter estimation schemes to try to home in on the appropriate value of ω by estimating the rate at which the iteration is converging. For coefficient matrices of a special class called consistently ordered with property A (see Young [215]), which includes certain orderings of matrices arising from the discretization of el- liptic PDEs, there is a direct relationship between the spectra of the Jacobi and SOR iteration
- 21. 2.2. STATIONARY ITERATIVE METHODS 11 matrices. In principle, given the spectral radius ρ of the Jacobi iteration matrix, one can determine a priori the theoretically optimal value of ω for SOR: ωopt = 2 1 + 1 − ρ2 . (2.8) This is seldom done, since calculating the spectral radius of the Jacobi matrix requires an imprac- tical amount of computation. However, relatively inexpensive rough estimates of ρ (for example, from the power method, see Golub and Van Loan [108, p. 351]) can yield reasonable estimates for the optimal value of ω. 2.2.4 The Symmetric Successive Overrelaxation Method If we assume that the coefficient matrix A is symmetric, then the Symmetric Successive Overre- laxation method, or SSOR, combines two SOR sweeps together in such a way that the resulting iteration matrix is similar to a symmetric matrix. Specifically, the first SOR sweep is carried out as in (2.7), but in the second sweep the unknowns are updated in the reverse order. That is, SSOR is a forward SOR sweep followed by a backward SOR sweep. The similarity of the SSOR iteration matrix to a symmetric matrix permits the application of SSOR as a preconditioner for other iter- ative schemes for symmetric matrices. Indeed, this is the primary motivation for SSOR since its convergence rate, with an optimal value of ω, is usually slower than the convergence rate of SOR with optimal ω (see Young [215, page 462]). For details on using SSOR as a preconditioner, see Chapter 3. In matrix terms, the SSOR iteration can be expressed as follows: x(k) = B1B2x(k−1) + ω(2 − ω)(D − ωU)−1 D(D − ωL)−1 b, (2.9) where B1 = (D − ωU)−1 (ωL + (1 − ω)D), and B2 = (D − ωL)−1 (ωU + (1 − ω)D). Note that B2 is simply the iteration matrix for SOR from (2.7), and that B1 is the same, but with the roles of L and U reversed. The pseudocode for the SSOR algorithm is given in Figure 2.4. 2.2.5 Notes and References The modern treatment of iterative methods dates back to the relaxation method of Southwell [192]. This was the precursor to the SOR method, though the order in which approximations to the un- knowns were relaxed varied during the computation. Specifically, the next unknown was chosen based upon estimates of the location of the largest error in the current approximation. Because of this, Southwell’s relaxation method was considered impractical for automated computing. It is interesting to note that the introduction of multiple-instruction, multiple data-stream (MIMD) paral- lel computers has rekindled an interest in so-called asynchronous, or chaotic iterative methods (see Chazan and Miranker [53], Baudet [30], and Elkin [91]), which are closely related to Southwell’s original relaxation method. In chaotic methods, the order of relaxation is unconstrained, thereby eliminating costly synchronization of the processors, though the effect on convergence is difficult to predict. The notion of accelerating the convergence of an iterative method by extrapolation predates the development of SOR. Indeed, Southwell used overrelaxation to accelerate the convergence of his original relaxation method. More recently, the ad hoc SOR method, in which a different
- 22. 12 CHAPTER 2. ITERATIVE METHODS Choose an initial guess x(0) to the solution x. Let x( 1 2 ) = x(0) . for k = 1, 2, . . . for i = 1, 2, . . . , n σ = 0 for j = 1, 2, . . . , i − 1 σ = σ + ai,jx (k− 1 2 ) j end for j = i + 1, . . . , n σ = σ + ai,jx (k−1) j end σ = (bi − σ)/ai,i x (k− 1 2 ) i = x (k−1) i + ω(σ − x (k−1) i ) end for i = n, n − 1, . . . , 1 σ = 0 for j = 1, 2, . . . , i − 1 σ = σ + ai,jx (k− 1 2 ) j end for j = i + 1, . . . , n σ = σ + ai,jx (k) j end x (k) i = x (k− 1 2 ) i + ω(σ − x (k− 1 2 ) i ) end check convergence; continue if necessary end Figure 2.4: The SSOR Method relaxation factor ω is used for updating each variable, has given impressive results for some classes of problems (see Ehrlich [82]). The three main references for the theory of stationary iterative methods are Varga [209], Young [215] and Hageman and Young [119]. The reader is referred to these books (and the references therein) for further details concerning the methods described in this section. 2.3 Nonstationary Iterative Methods Nonstationary methods differ from stationary methods in that the computations involve informa- tion that changes at each iteration. Typically, constants are computed by taking inner products of residuals or other vectors arising from the iterative method. 2.3.1 Conjugate Gradient Method (CG) The Conjugate Gradient method is an effective method for symmetric positive definite systems. It is the oldest and best known of the nonstationary methods discussed here. The method proceeds by generating vector sequences of iterates (i.e., successsive approximations to the solution), residuals
- 23. 2.3. NONSTATIONARY ITERATIVE METHODS 13 Compute r(0) = b − Ax(0) for some initial guess x(0) for i = 1, 2, . . . solve Mz(i−1) = r(i−1) ρi−1 = r(i−1)T z(i−1) if i = 1 p(1) = z(0) else βi−1 = ρi−1/ρi−2 p(i) = z(i−1) + βi−1p(i−1) endif q(i) = Ap(i) αi = ρi−1/p(i)T q(i) x(i) = x(i−1) + αip(i) r(i) = r(i−1) − αiq(i) check convergence; continue if necessary end Figure 2.5: The Preconditioned Conjugate Gradient Method corresponding to the iterates, and search directions used in updating the iterates and residuals. Although the length of these sequences can become large, only a small number of vectors needs to be kept in memory. In every iteration of the method, two inner products are performed in order to compute update scalars that are defined to make the sequences satisfy certain orthogonality conditions. On a symmetric positive definite linear system these conditions imply that the distance to the true solution is minimized in some norm. The iterates x(i) are updated in each iteration by a multiple (αi) of the search direction vec- tor p(i) : x(i) = x(i−1) + αip(i) . Correspondingly the residuals r(i) = b − Ax(i) are updated as r(i) = r(i−1) − αq(i) where q(i) = Ap(i) . (2.10) The choice α = αi = r(i−1)T r(i−1) /p(i)T Ap(i) minimizes r(i)T A−1 r(i) over all possible choices for α in equation (2.10). The search directions are updated using the residuals p(i) = r(i) + βi−1p(i−1) , (2.11) where the choice βi = r(i)T r(i) /r(i−1)T r(i−1) ensures that p(i) and Ap(i−1) – or equivalently, r(i) and r(i−1) – are orthogonal. In fact, one can show that this choice of βi makes p(i) and r(i) orthogonal to all previous Ap(j) and r(j) respectively. The pseudocode for the Preconditioned Conjugate Gradient Method is given in Figure 2.5. It uses a preconditioner M; for M = I one obtains the unpreconditioned version of the Conjugate Gradient Algorithm. In that case the algorithm may be further simplified by skipping the “solve” line, and replacing z(i−1) by r(i−1) (and z(0) by r(0) ). Theory The unpreconditioned conjugate gradient method constructs the ith iterate x(i) as an element of x(0) + span{r(0) , . . . , Ai−1 r(0) } so that (x(i) − ˆx)T A(x(i) − ˆx) is minimized, where ˆx is the exact
- 24. 14 CHAPTER 2. ITERATIVE METHODS solution of Ax = b. This minimum is guaranteed to exist in general only if A is symmetric positive definite. The preconditioned version of the method uses a different subspace for constructing the iterates, but it satisfies the same minimization property, although over this different subspace. It requires in addition that the preconditioner M is symmetric and positive definite. The above minimization of the error is equivalent to the residuals r(i) = b − Ax(i) being M−1 orthogonal (that is, r(i)T M−1 r(j) = 0 if i = j). Since for symmetric A an orthogonal basis for the Krylov subspace span{r(0) , . . . , Ai−1 r(0) } can be constructed with only three-term recurrences, such a recurrence also suffices for generating the residuals. In the Conjugate Gradient method two coupled two-term recurrences are used; one that updates residuals using a search direction vector, and one updating the search direction with a newly computed residual. This makes the Conjugate Gradient Method quite attractive computationally. There is a close relationship between the Conjugate Gradient method and the Lanczos method for determining eigensystems, since both are based on the construction of an orthogonal basis for the Krylov subspace, and a similarity transformation of the coefficient matrix to tridiagonal form. The coefficients computed during the CG iteration then arise from the LU factorization of this tridiagonal matrix. From the CG iteration one can reconstruct the Lanczos process, and vice versa; see Paige and Saunders [167] and Golub and Van Loan [108, §10.2.6]. This relationship can be exploited to obtain relevant information about the eigensystem of the (preconditioned) matrix A; see §5.1. Convergence Accurate predictions of the convergence of iterative methods are difficult to make, but useful bounds can often be obtained. For the Conjugate Gradient method, the error can be bounded in terms of the spectral condition number κ2 of the matrix M−1 A. (Recall that if λmax and λmin are the largest and smallest eigenvalues of a symmetric positive definite matrix B, then the spectral condition number of B is κ2(B) = λmax(B)/λmin(B)). If ˆx is the exact solution of the linear system Ax = b, with symmetric positive definite matrix A, then for CG with symmetric positive definite preconditioner M, it can be shown that x(i) − ˆx A ≤ 2αi x(0) − ˆx A (2.12) where α = ( √ κ2 − 1)/( √ κ2 + 1) (see Golub and Van Loan [108, §10.2.8], and Kaniel [130]), and y 2 A ≡ (y, Ay). From this relation we see that the number of iterations to reach a relative reduction of in the error is proportional to √ κ2. In some cases, practical application of the above error bound is straightforward. For example, elliptic second order partial differential equations typically give rise to coefficient matrices A with κ2(A) = O(h−2 ) (where h is the discretization mesh width), independent of the order of the finite elements or differences used, and of the number of space dimensions of the problem (see for instance Axelsson and Barker [14, §5.5]). Thus, without preconditioning, we expect a number of iterations proportional to h−1 for the Conjugate Gradient method. Other results concerning the behavior of the Conjugate Gradient algorithm have been obtained. If the extremal eigenvalues of the matrix M−1 A are well separated, then one often observes so- called superlinear convergence (see Concus, Golub and O’Leary [57]); that is, convergence at a rate that increases per iteration. This phenomenon is explained by the fact that CG tends to elimi- nate components of the error in the direction of eigenvectors associated with extremal eigenvalues first. After these have been eliminated, the method proceeds as if these eigenvalues did not exist in the given system, i.e., the convergence rate depends on a reduced system with a (much) smaller condition number (for an analysis of this, see Van der Sluis and Van der Vorst [197]). The effective- ness of the preconditioner in reducing the condition number and in separating extremal eigenvalues can be deduced by studying the approximated eigenvalues of the related Lanczos process.
- 25. 2.3. NONSTATIONARY ITERATIVE METHODS 15 Implementation The Conjugate Gradient method involves one matrix-vector product, three vector updates, and two inner products per iteration. Some slight computational variants exist that have the same structure (see Reid [178]). Variants that cluster the inner products, a favorable property on parallel machines, are discussed in §4.4. For a discussion of the Conjugate Gradient method on vector and shared memory computers, see Dongarra, et al. [70, 165]. For discussions of the method for more general parallel architectures see Demmel, Heath and Van der Vorst [66] and Ortega [165], and the references therein. Further references A more formal presentation of CG, as well as many theoretical properties, can be found in the textbook by Hackbusch [117]. Shorter presentations are given in Axelsson and Barker [14] and Golub and Van Loan [108]. An overview of papers published in the first 25 years of existence of the method is given in Golub and O’Leary [107]. 2.3.2 MINRES and SYMMLQ The Conjugate Gradient method can be viewed as a special variant of the Lanczos method (see §5.1) for positive definite symmetric systems. The MINRES and SYMMLQ methods are variants that can be applied to symmetric indefinite systems. The vector sequences in the Conjugate Gradient method correspond to a factorization of a tridi- agonal matrix similar to the coefficient matrix. Therefore, a breakdown of the algorithm can occur corresponding to a zero pivot if the matrix is indefinite. Furthermore, for indefinite matrices the minimization property of the Conjugate Gradient method is no longer well-defined. The MINRES and SYMMLQ methods are variants of the CG method that avoid the LU-factorization and do not suffer from breakdown. MINRES minimizes the residual in the 2-norm. SYMMLQ solves the projected system, but does not minimize anything (it keeps the residual orthogonal to all previous ones). The convergence behavior of Conjugate Gradients and MINRES for indefinite systems was analyzed by Paige, Parlett, and Van der Vorst [166]. Theory When A is not positive definite, but symmetric, we can still construct an orthogonal basis for the Krylov subspace by three term recurrence relations. Eliminating the search directions in equa- tions (2.10) and (2.11) gives a recurrence Ar(i) = r(i+1) ti+1,i + r(i) ti,i + r(i−1) ti−1,i, which can be written in matrix form as ARi = Ri+1 ¯Ti, where ¯Ti is an (i + 1) × i tridiagonal matrix ¯Ti = ← i → ... ... ... ... ... ... ... ... ... ... ... . ↑ i + 1 ↓
- 26. 16 CHAPTER 2. ITERATIVE METHODS In this case we have the problem that (·, ·)A no longer defines an inner product. However we can still try to minimize the residual in the 2-norm by obtaining x(i) ∈ {r(0) , Ar(0) , . . . , Ai−1 r(0) }, x(i) = Ri ¯y that minimizes Ax(i) − b 2 = ARi ¯y − b 2 = Ri+1 ¯Tiy − b 2. Now we exploit the fact that if Di+1 ≡ diag( r(0) 2, r(1) 2, ..., r(i) 2), then Ri+1D−1 i+1 is an orthonormal transformation with respect to the current Krylov subspace: Ax(i) − b 2 = Di+1 ¯Tiy − r(0) 2e(1) 2 and this final expression can simply be seen as a minimum norm least squares problem. The element in the (i + 1, i) position of ¯Ti can be annihilated by a simple Givens rotation and the resulting upper bidiagonal system (the other subdiagonal elements having been removed in previous iteration steps) can simply be solved, which leads to the MINRES method (see Paige and Saunders [167]). Another possibility is to solve the system Tiy = r(0) 2e(1) , as in the CG method (Ti is the upper i × i part of ¯Ti). Other than in CG we cannot rely on the existence of a Choleski decomposition (since A is not positive definite). An alternative is then to decompose Ti by an LQ-decomposition. This again leads to simple recurrences and the resulting method is known as SYMMLQ (see Paige and Saunders [167]). 2.3.3 CG on the Normal Equations, CGNE and CGNR The CGNE and CGNR methods are the simplest methods for nonsymmetric or indefinite systems. Since other methods for such systems are in general rather more complicated than the Conjugate Gradient method, transforming the system to a symmetric definite one and then applying the Con- jugate Gradient method is attractive for its coding simplicity. Theory If a system of linear equations Ax = b has a nonsymmetric, possibly indefinite (but nonsingular), coefficient matrix, one obvious attempt at a solution is to apply Conjugate Gradient to a related symmetric positive definite system, AT Ax = AT b. While this approach is easy to understand and code, the convergence speed of the Conjugate Gradient method now depends on the square of the condition number of the original coefficient matrix. Thus the rate of convergence of the CG procedure on the normal equations may be slow. Several proposals have been made to improve the numerical stability of this method. The best known is by Paige and Saunders [168] and is based upon applying the Lanczos method to the auxiliary system I A AT 0 r x = b 0 . A clever execution of this scheme delivers the factors L and U of the LU-decomposition of the tridiagonal matrix that would have been computed by carrying out the Lanczos procedure with AT A. Another means for improving the numerical stability of this normal equations approach is sug- gested by Bj¨orck and Elfving in [34]. The observation that the matrix AT A is used in the construc- tion of the iteration coefficients through an inner product like (p, AT Ap) leads to the suggestion that such an inner product be replaced by (Ap, Ap).
- 27. 2.3. NONSTATIONARY ITERATIVE METHODS 17 2.3.4 Generalized Minimal Residual (GMRES) The Generalized Minimal Residual method is an extension of MINRES (which is only applicable to symmetric systems) to unsymmetric systems. Like MINRES, it generates a sequence of orthogonal vectors, but in the absence of symmetry this can no longer be done with short recurrences; instead, all previously computed vectors in the orthogonal sequence have to be retained. For this reason, “restarted” versions of the method are used. In the Conjugate Gradient method, the residuals form an orthogonal basis for the space span{r(0) , Ar(0) , A2 r(0) , . . .}. In GMRES, this basis is formed explicitly: w(i) = Av(i) for k = 1, . . . , i w(i) = w(i) − (w(i) , v(k) )v(k) end v(i+1) = w(i) / w(i) The reader may recognize this as a modified Gram-Schmidt orthogonalization. Applied to the Krylov sequence {Ak r(0) } this orthogonalization is called the “Arnoldi method” [6]. The inner product coefficients (w(i) , v(k) ) and w(i) are stored in an upper Hessenberg matrix. The GMRES iterates are constructed as x(i) = x(0) + y1v(1) + · · · + yiv(i) , where the coefficients yk have been chosen to minimize the residual norm b − Ax(i) . The GMRES algorithm has the property that this residual norm can be computed without the iterate having been formed. Thus, the expensive action of forming the iterate can be postponed until the residual norm is deemed small enough. The pseudocode for the restarted GMRES(m) algorithm with preconditioner M is given in Figure 2.6. Theory The Generalized Minimum Residual (GMRES) method is designed to solve nonsymmetric linear systems (see Saad and Schultz [188]). The most popular form of GMRES is based on the modified Gram-Schmidt procedure, and uses restarts to control storage requirements. If no restarts are used, GMRES (like any orthogonalizing Krylov-subspace method) will con- verge in no more than n steps (assuming exact arithmetic). Of course this is of no practical value when n is large; moreover, the storage and computational requirements in the absence of restarts are prohibitive. Indeed, the crucial element for successful application of GMRES(m) revolves around the decision of when to restart; that is, the choice of m. Unfortunately, there exist examples for which the method stagnates and convergence takes place only at the nth step. For such systems, any choice of m less than n fails to converge. Saad and Schultz [188] have proven several useful results. In particular, they show that if the coefficient matrix A is real and nearly positive definite, then a “reasonable” value for m may be selected. Implications of the choice of m are discussed below. Implementation A common implementation of GMRES is suggested by Saad and Schultz in [188] and relies on using modified Gram-Schmidt orthogonalization. Householder transformations, which are rela- tively costly but stable, have also been proposed. The Householder approach results in a three-fold increase in work; however, convergence may be better, especially for ill-conditioned systems (see Walker [212]). From the point of view of parallelism, Gram-Schmidt orthogonalization may be
- 28. 18 CHAPTER 2. ITERATIVE METHODS x(0) is an initial guess for j = 1, 2, .... Solve r from Mr = b − Ax(0) v(1) = r/ r 2 s := r 2e1 for i = 1, 2, ..., m Solve w from Mw = Av(i) for k = 1, ..., i hk,i = (w, v(k) ) w = w − hk,iv(k) end hi+1,i = w 2 v(i+1) = w/hi+1,i apply J1, ..., Ji−1 on (h1,i, ..., hi+1,i) construct Ji, acting on ith and (i + 1)st component of h.,i, such that (i + 1)st component of Jih.,i is 0 s := Jis if s(i + 1) is small enough then (UPDATE(˜x, i) and quit) end UPDATE(˜x, m) end In this scheme UPDATE(˜x, i) replaces the following computations: Compute y as the solution of Hy = ˜s, in which the upper i × i triangular part of H has hi,j as its elements (in least squares sense if H is singular), ˜s represents the first i components of s ˜x = x(0) + y1v(1) + y2v(2) + ... + yiv(i) s(i+1) = b − A˜x 2 if ˜x is an accurate enough approximation then quit else x(0) = ˜x Figure 2.6: The Preconditioned GMRES(m) Method preferred, giving up some stability for better parallelization properties (see Demmel, Heath and Van der Vorst [66]). Here we adopt the Modified Gram-Schmidt approach. The major drawback to GMRES is that the amount of work and storage required per iteration rises linearly with the iteration count. Unless one is fortunate enough to obtain extremely fast con- vergence, the cost will rapidly become prohibitive. The usual way to overcome this limitation is by restarting the iteration. After a chosen number (m) of iterations, the accumulated data are cleared and the intermediate results are used as the initial data for the next m iterations. This procedure is repeated until convergence is achieved. The difficulty is in choosing an appropriate value for m. If m is “too small”, GMRES(m) may be slow to converge, or fail to converge entirely. A value of m that is larger than necessary involves excessive work (and uses more storage). Unfortunately, there are no definite rules governing the choice of m—choosing when to restart is a matter of experience. For a discussion of GMRES for vector and shared memory computers see Dongarra et al. [70];
- 29. 2.3. NONSTATIONARY ITERATIVE METHODS 19 for more general architectures, see Demmel, Heath and Van der Vorst [66]. 2.3.5 BiConjugate Gradient (BiCG) The Conjugate Gradient method is not suitable for nonsymmetric systems because the residual vectors cannot be made orthogonal with short recurrences (for proof of this see Voevodin [211] or Faber and Manteuffel [95]). The GMRES method retains orthogonality of the residuals by using long recurrences, at the cost of a larger storage demand. The BiConjugate Gradient method takes another approach, replacing the orthogonal sequence of residuals by two mutually orthogonal sequences, at the price of no longer providing a minimization. The update relations for residuals in the Conjugate Gradient method are augmented in the BiConjugate Gradient method by relations that are similar but based on AT instead of A. Thus we update two sequences of residuals r(i) = r(i−1) − αiAp(i) , ˜r(i) = ˜r(i−1) − αiAT ˜p(i) , and two sequences of search directions p(i) = r(i−1) + βi−1p(i−1) , ˜p(i) = ˜r(i−1) + βi−1 ˜p(i−1) . The choices αi = ˜r(i−1)T r(i−1) ˜p(i)T Ap(i) , βi = ˜r(i)T r(i) ˜r(i−1)T r(i−1) ensure the bi-orthogonality relations ˜r(i)T r(j) = ˜p(i)T Ap(j) = 0 if i = j. The pseudocode for the Preconditioned BiConjugate Gradient Method with preconditioner M is given in Figure 2.7. Convergence Few theoretical results are known about the convergence of BiCG. For symmetric positive defi- nite systems the method delivers the same results as CG, but at twice the cost per iteration. For nonsymmetric matrices it has been shown that in phases of the process where there is significant reduction of the norm of the residual, the method is more or less comparable to full GMRES (in terms of numbers of iterations) (see Freund and Nachtigal [101]). In practice this is often con- firmed, but it is also observed that the convergence behavior may be quite irregular, and the method may even break down. The breakdown situation due to the possible event that z(i−1)T ˜r(i−1) ≈ 0 can be circumvented by so-called look-ahead strategies (see Parlett, Taylor and Liu [171]). This leads to complicated codes and is beyond the scope of this book. The other breakdown situation, ˜p(i)T q(i) ≈ 0, occurs when the LU-decomposition fails (see the theory subsection of §2.3.1), and can be repaired by using another decomposition. This is done in QMR (see §2.3.6). Sometimes, breakdown or near-breakdown situations can be satisfactorily avoided by a restart at the iteration step immediately before the (near-) breakdown step. Another possibility is to switch to a more robust (but possibly more expensive) method, like GMRES. Implementation BiCG requires computing a matrix-vector product Ap(k) and a transpose product AT ˜p(k) . In some applications the latter product may be impossible to perform, for instance if the matrix is not formed explicitly and the regular product is only given in operation form, for instance as a function call evaluation.
- 30. 20 CHAPTER 2. ITERATIVE METHODS Compute r(0) = b − Ax(0) for some initial guess x(0) . Choose ˜r(0) (for example, ˜r(0) = r(0) ). for i = 1, 2, . . . solve Mz(i−1) = r(i−1) solve MT ˜z(i−1) = ˜r(i−1) ρi−1 = z(i−1)T ˜r(i−1) if ρi−1 = 0, method fails if i = 1 p(i) = z(i−1) ˜p(i) = ˜z(i−1) else βi−1 = ρi−1/ρi−2 p(i) = z(i−1) + βi−1p(i−1) ˜p(i) = ˜z(i−1) + βi−1 ˜p(i−1) endif q(i) = Ap(i) ˜q(i) = AT ˜p(i) αi = ρi−1/˜p(i)T q(i) x(i) = x(i−1) + αip(i) r(i) = r(i−1) − αiq(i) ˜r(i) = ˜r(i−1) − αi ˜q(i) check convergence; continue if necessary end Figure 2.7: The Preconditioned BiConjugate Gradient Method In a parallel environment, the two matrix-vector products can theoretically be performed si- multaneously; however, in a distributed-memory environment, there will be extra communication costs associated with one of the two matrix-vector products, depending upon the storage scheme for A. A duplicate copy of the matrix will alleviate this problem, at the cost of doubling the storage requirements for the matrix. Care must also be exercised in choosing the preconditioner, since similar problems arise during the two solves involving the preconditioning matrix. It is difficult to make a fair comparison between GMRES and BiCG. GMRES really minimizes a residual, but at the cost of increasing work for keeping all residuals orthogonal and increasing demands for memory space. BiCG does not minimize a residual, but often its accuracy is com- parable to GMRES, at the cost of twice the amount of matrix vector products per iteration step. However, the generation of the basis vectors is relatively cheap and the memory requirements are modest. Several variants of BiCG have been proposed that increase the effectiveness of this class of methods in certain circumstances. These variants (CGS and Bi-CGSTAB) will be discussed in coming subsections. 2.3.6 Quasi-Minimal Residual (QMR) The BiConjugate Gradient method often displays rather irregular convergence behavior. Moreover, the implicit LU decomposition of the reduced tridiagonal system may not exist, resulting in break- down of the algorithm. A related algorithm, the Quasi-Minimal Residual method of Freund and Nachtigal [101], [102] attempts to overcome these problems. The main idea behind this algorithm
- 31. 2.3. NONSTATIONARY ITERATIVE METHODS 21 is to solve the reduced tridiagonal system in a least squares sense, similar to the approach fol- lowed in GMRES. Since the constructed basis for the Krylov subspace is bi-orthogonal, rather than orthogonal as in GMRES, the obtained solution is viewed as a quasi-minimal residual solution, which explains the name. Additionally, QMR uses look-ahead techniques to avoid breakdowns in the underlying Lanczos process, which makes it more robust than BiCG. Convergence The convergence behavior of QMR is typically much smoother than for BiCG. Freund and Nachti- gal [101] present quite general error bounds which show that QMR may be expected to converge about as fast as GMRES. From a relation between the residuals in BiCG and QMR (Freund and Nachtigal [101, relation (5.10)]) one may deduce that at phases in the iteration process where BiCG makes significant progress, QMR has arrived at about the same approximation for ˆx. On the other hand, when BiCG makes no progress at all, QMR may still show slow convergence. The look-ahead steps in the QMR method prevent breakdown in all cases but the so-called “incurable breakdown”, where no number of look-ahead steps would yield a next iterate. Implementation The pseudocode for the Preconditioned Quasi Minimal Residual Method, with preconditioner M = M1M2, is given in Figure 2.8. This algorithm follows the two term recurrence version without look-ahead, presented by Freund and Nachtigal [102] as Algorithm 7.1. This version of QMR is simpler to implement than the full QMR method with look-ahead, but it is susceptible to breakdown of the underlying Lanczos process. (Other implementational variations are whether to scale Lanczos vectors or not, or to use three-term recurrences instead of coupled two-term re- currences. Such decisions usually have implications for the stability and the efficiency of the algo- rithm.) A professional implementation of QMR with look-ahead is given in Freund and Nachtigal’s QMRPACK, which is available through netlib; see Appendix A. We have modified the algorithm to include a relatively inexpensive recurrence relation for the computation of the residual vector. This requires a few extra vectors of storage and vector update operations per iteration, but it avoids expending a matrix-vector product on the residual calculation. Also, the algorithm has been modified so that only two full preconditioning steps are required instead of three. Computation of the residual is done for the convergence test. If one uses right (or post) precon- ditioning, that is M1 = I, then a cheap upper bound for r(i) can be computed in each iteration, avoiding the recursions for r(i) . For details, see Freund and Nachtigal [101, proposition 4.1]. This upper bound may be pessimistic by a factor of at most √ i + 1. QMR has roughly the same problems with respect to vector and parallel implementation as BiCG. The scalar overhead per iteration is slightly more than for BiCG. In all cases where the slightly cheaper BiCG method converges irregularly (but fast enough), there seems little reason to avoid QMR. 2.3.7 Conjugate Gradient Squared Method (CGS) In BiCG, the residual vector r(i) can be regarded as the product of r(0) and an ith degree polynomial in A, that is r(i) = Pi(A)r(0) . (2.13) This same polynomial satisfies ˜r(i) = Pi(A)˜r(0) so that ρi = (˜r(i) , r(i) ) = (Pi(AT )˜r(0) , Pi(A)r(0) ) = (˜r(i) , P2 i (A)r(0) ). (2.14)
- 32. 22 CHAPTER 2. ITERATIVE METHODS Compute r(0) = b − Ax(0) for some initial guess x(0) ˜v(1) = r(0) ; solve M1y = ˜v(1) ; ρ1 = y 2 Choose ˜w(1) , for example ˜w(1) = r(0) solve Mt 2z = ˜w(1) ; ξ1 = z 2 γ0 = 1; η0 = −1 for i = 1, 2, . . . if ρi = 0 or ξi = 0 method fails v(i) = ˜v(i) /ρi; y = y/ρi w(i) = ˜w(i) /ξi; z = z/ξi δi = zT y; if δi = 0 method fails solve M2 ˜y = y solve MT 1 ˜z = z if i = 1 p(1) = ˜y; q(1) = ˜z else p(i) = ˜y − (ξiδi/ i−1)p(i−1) q(i) = ˜z − (ρiδi/ i−1)q(i−1) endif ˜p = Ap(i) i = q(i)T ˜p; if i = 0 method fails βi = i/δi; if βi = 0 method fails ˜v(i+1) = ˜p − βiv(i) solve M1y = ˜v(i+1) ρi+1 = y 2 ˜w(i+1) = AT q(i) − βiw(i) solve MT 2 z = ˜w(i+1) ξi+1 = z 2 θi = ρi+1/(γi−1|βi|); γi = 1/ 1 + θ2 i ; if γi = 0 method fails ηi = −ηi−1ρiγ2 i /(βiγ2 i−1) if i = 1 d(1) = η1p(1) ; s(1) = η1 ˜p else d(i) = ηip(i) + (θi−1γi)2 d(i−1) s(i) = ηi ˜p + (θi−1γi)2 s(i−1) endif x(i) = x(i−1) + d(i) r(i) = r(i−1) − s(i) check convergence; continue if necessary end Figure 2.8: The Preconditioned Quasi Minimal Residual Method without Look-ahead
- 33. 2.3. NONSTATIONARY ITERATIVE METHODS 23 Compute r(0) = b − Ax(0) for some initial guess x(0) Choose ˜r (for example, ˜r = r(0) ) for i = 1, 2, . . . ρi−1 = ˜rT r(i−1) if ρi−1 = 0 method fails if i = 1 u(1) = r(0) p(1) = u(1) else βi−1 = ρi−1/ρi−2 u(i) = r(i−1) + βi−1q(i−1) p(i) = u(i) + βi−1(q(i−1) + βi−1p(i−1) ) endif solve M ˆp = p(i) ˆv = Aˆp αi = ρi−1/˜rT ˆv q(i) = u(i) − αiˆv solve M ˆu = u(i) + q(i) x(i) = x(i−1) + αi ˆu ˆq = Aˆu r(i) = r(i−1) − αi ˆq check convergence; continue if necessary end Figure 2.9: The Preconditioned Conjugate Gradient Squared Method This suggests that if Pi(A) reduces r(0) to a smaller vector r(i) , then it might be advantageous to apply this “contraction” operator twice, and compute P2 i (A)r(0) . Equation (2.14) shows that the iteration coefficients can still be recovered from these vectors, and it turns out to be easy to find the corresponding approximations for x. This approach leads to the Conjugate Gradient Squared method (see Sonneveld [191]). Convergence Often one observes a speed of convergence for CGS that is about twice as fast as for BiCG, which is in agreement with the observation that the same “contraction” operator is applied twice. However, there is no reason that the “contraction” operator, even if it really reduces the initial residual r(0) , should also reduce the once reduced vector r(k) = Pk(A)r(0) . This is evidenced by the often highly irregular convergence behavior of CGS. One should be aware of the fact that local corrections to the current solution may be so large that cancellation effects occur. This may lead to a less accurate solution than suggested by the updated residual (see Van der Vorst [205]). The method tends to diverge if the starting guess is close to the solution. Implementation CGS requires about the same number of operations per iteration as BiCG, but does not involve computations with AT . Hence, in circumstances where computation with AT is impractical, CGS may be attractive.
- 34. 24 CHAPTER 2. ITERATIVE METHODS Compute r(0) = b − Ax(0) for some initial guess x(0) Choose ˜r (for example, ˜r = r(0) ) for i = 1, 2, . . . ρi−1 = ˜rT r(i−1) if ρi−1 = 0 method fails if i = 1 p(i) = r(i−1) else βi−1 = (ρi−1/ρi−2)(αi−1/ωi−1) p(i) = r(i−1) + βi−1(p(i−1) − ωi−1v(i−1) ) endif solve M ˆp = p(i) v(i) = Aˆp αi = ρi−1/˜rT v(i) s = r(i−1) − αiv(i) check norm of s; if small enough: set x(i) = x(i−1) + αi ˆp and stop solve Mˆs = s t = Aˆs ωi = tT s/tT t x(i) = x(i−1) + αi ˆp + ωiˆs r(i) = s − ωit check convergence; continue if necessary for continuation it is necessary that ωi = 0 end Figure 2.10: The Preconditioned BiConjugate Gradient Stabilized Method The pseudocode for the Preconditioned Conjugate Gradient Squared Method with precondi- tioner M is given in Figure 2.9. 2.3.8 BiConjugate Gradient Stabilized (Bi-CGSTAB) The BiConjugate Gradient Stabilized method (Bi-CGSTAB) was developed to solve nonsymmet- ric linear systems while avoiding the often irregular convergence patterns of the Conjugate Gra- dient Squared method (see Van der Vorst [205]). Instead of computing the CGS sequence i → P2 i (A)r(0) , Bi-CGSTAB computes i → Qi(A)Pi(A)r(0) where Qi is an ith degree polynomial describing a steepest descent update. Convergence Bi-CGSTAB often converges about as fast as CGS, sometimes faster and sometimes not. CGS can be viewed as a method in which the BiCG “contraction” operator is applied twice. Bi-CGSTAB can be interpreted as the product of BiCG and repeatedly applied GMRES(1). At least locally, a residual vector is minimized, which leads to a considerably smoother convergence behavior. On the other hand, if the local GMRES(1) step stagnates, then the Krylov subspace is not expanded, and Bi-CGSTAB will break down. This is a breakdown situation that can occur in addition to the other breakdown possiblities in the underlying BiCG algorithm. This type of breakdown may be avoided by combining BiCG with other methods, i.e., by selecting other values for ωi (see the
- 35. 2.3. NONSTATIONARY ITERATIVE METHODS 25 algorithm). One such alternative is Bi-CGSTAB2 (see Gutknecht [114]); more general approaches are suggested by Sleijpen and Fokkema in [189]. Note that Bi-CGSTAB has two stopping tests: if the method has already converged at the first test on the norm of s, the subsequent update would be numerically questionable. Additionally, stopping on the first test saves a few unnecessary operations, but this is of minor importance. Implementation Bi-CGSTAB requires two matrix-vector products and four inner products, i.e., two inner products more than BiCG and CGS. The pseudocode for the Preconditioned BiConjugate Gradient Stabilized Method with precon- ditioner M is given in Figure 2.10. 2.3.9 Chebyshev Iteration Chebyshev Iteration is another method for solving nonsymmetric problems (see Golub and Van Loan [108, §10.1.5] and Varga [209, Chapter 5]). Chebyshev Iteration avoids the computation of inner products as is necessary for the other nonstationary methods. For some distributed memory architectures these inner products are a bottleneck with respect to efficiency. The price one pays for avoiding inner products is that the method requires enough knowledge about the spectrum of the coefficient matrix A that an ellipse enveloping the spectrum can be identified; however this difficulty can be overcome via an adaptive construction developed by Manteuffel [145], and implemented by Ashby [7]. Chebyshev iteration is suitable for any nonsymmetric linear system for which the enveloping ellipse does not include the origin. Comparison with other methods Comparing the pseudocode for Chebyshev Iteration with the pseudocode for the Conjugate Gra- dient method shows a high degree of similarity, except that no inner products are computed in Chebyshev Iteration. Scalars c and d must be selected so that they define a family of ellipses with common center d > 0 and foci d + c and d − c which contain the ellipse that encloses the spectrum (or more general, field of values) of A and for which the rate r of convergence is minimal: r = a + √ a2 − c2 d + √ d2 − c2 , (2.15) where a is the length of the x-axis of the ellipse. We provide code in which it is assumed that c and d are known. For code including the adap- tive detemination of these iteration parameters the reader is referred to Ashby [7]. The Chebyshev method has the advantage over GMRES that only short recurrences are used. On the other hand, GMRES is guaranteed to generate the smallest residual over the current search space. The BiCG methods, which also use short recurrences, do not minimize the residual in a suitable norm; how- ever, unlike Chebyshev iteration, they do not require estimation of parameters (the spectrum of A). Finally, GMRES and BiCG may be more effective in practice, because of superlinear convergence behavior, which cannot be expected for Chebyshev. For symmetric positive definite systems the “ellipse” enveloping the spectrum degenerates to the interval [λmin, λmax] on the positive x-axis, where λmin and λmax are the smallest and largest eigenvalues of M−1 A. In circumstances where the computation of inner products is a bottleneck, it may be advantageous to start with CG, compute estimates of the extremal eigenvalues from the CG coefficients, and then after sufficient convergence of these approximations switch to Chebyshev Iteration. A similar strategy may be adopted for a switch from GMRES, or BiCG-type methods, to Chebyshev Iteration.
- 36. 26 CHAPTER 2. ITERATIVE METHODS Compute r(0) = b − Ax(0) for some initial guess x(0) . d = (λmax + λmin)/2, c = (λmax − λmin)/2. for i = 1, 2, . . . solve Mz(i−1) = r(i) . if i = 1 p(1) = z(0) α1 = 2/d else βi−1 = (cαi−1/2)2 αi = 1/(d − βi−1) p(i) = z(i−1) + βi−1p(i−1) . endif x(i) = x(i−1) + αip(i) . r(i) = b − Ax(i) (= r(i−1) − αiAp(i) ). check convergence; continue if necessary end Figure 2.11: The Preconditioned Chebyshev Method Convergence In the symmetric case (where A and the preconditioner M are both symmetric) for the Chebyshev Iteration we have the same upper bound as for the Conjugate Gradient method, provided c and d are computed from λmin and λmax (the extremal eigenvalues of the preconditioned matrix M−1 A). There is a severe penalty for overestimating or underestimating the field of values. For ex- ample, if in the symmetric case λmax is underestimated, then the method may diverge; if it is overestimated then the result may be very slow convergence. Similar statements can be made for the nonsymmetric case. This implies that one needs fairly accurate bounds on the spectrum of M−1 A for the method to be effective (in comparison with CG or GMRES). Implementation In Chebyshev Iteration the iteration parameters are known as soon as one knows the ellipse con- taining the eigenvalues (or rather, the field of values) of the operator. Therefore the computation of inner products, as is necessary in methods like GMRES or CG, is avoided. This avoids the synchronization points required of CG-type methods, so machines with hierarchical or distributed memory may achieve higher performance (it also suggests strong parallelization properties; for a discussion of this see Saad [184], and Dongarra, et al. [70]). Specifically, as soon as some segment of w is computed, we may begin computing, in sequence, corresponding segments of p, x, and r. The pseudocode for the Preconditioned Chebyshev Method with preconditioner M is given in Figure 2.11. It handles the case of a symmetric positive definite coefficient matrix A. The eigenvalues of M−1 A are assumed to be all real and in the interval [λmin, λmax], which does not include zero. 2.4 Computational Aspects of the Methods Efficient solution of a linear system is largely a function of the proper choice of iterative method. However, to obtain good performance, consideration must also be given to the computational ker-
- 37. 2.4. COMPUTATIONAL ASPECTS OF THE METHODS 27 Matrix- Inner Vector Precond Method Product SAXPY Product Solve JACOBI 1a GS 1 1a SOR 1 1a CG 2 3 1 1 GMRES i + 1 i + 1 1 1 BiCG 2 5 1/1 1/1 QMR 2 8+4bc 1/1 1/1 CGS 2 6 2 2 Bi-CGSTAB 4 6 2 2 CHEBYSHEV 2 1 1 Table 2.1: Summary of Operations for Iteration i. “a/b” means “a” multiplications with the matrix and “b” with its transpose. aThis method performs no real matrix vector product or preconditioner solve, but the number of operations is equivalent to a matrix-vector multiply. bTrue SAXPY operations + vector scalings. cLess for implementations that do not recursively update the residual. nels of the method and how efficiently they can be executed on the target architecture. This point is of particular importance on parallel architectures; see §4.4. Iterative methods are very different from direct methods in this respect. The performance of direct methods, both for dense and sparse systems, is largely that of the factorization of the matrix. This operation is absent in iterative methods (although preconditioners may require a setup phase), and with it, iterative methods lack dense matrix suboperations. Since such operations can be exe- cuted at very high efficiency on most current computer architectures, we expect a lower flop rate for iterative than for direct methods. (Dongarra and Van der Vorst [73] give some experimental results about this, and provide a benchmark code for iterative solvers.) Furthermore, the basic operations in iterative methods often use indirect addressing, depending on the data structure. Such operations also have a relatively low efficiency of execution. However, this lower efficiency of execution does not imply anything about the total solution time for a given system. Furthermore, iterative methods are usually simpler to implement than direct methods, and since no full factorization has to be stored, they can handle much larger sytems than direct methods. In this section we summarize for each method • Matrix properties. Not every method will work on every problem type, so knowledge of matrix properties is the main criterion for selecting an iterative method. • Computational kernels. Different methods involve different kernels, and depending on the problem or target computer architecture this may rule out certain methods. Table 2.2 lists the storage required for each method (without preconditioning). Note that we are not including the original system Ax = b and we ignore scalar storage. 1. Jacobi Method • Extremely easy to use, but unless the matrix is “strongly” diagonally dominant, this method is probably best only considered as an introduction to iterative methods or as a preconditioner in a nonstationary method. • Trivial to parallelize.
- 38. 28 CHAPTER 2. ITERATIVE METHODS Method Storage Reqmts JACOBI matrix + 3n SOR matrix + 2n CG matrix + 6n GMRES matrix + (i + 5)n BiCG matrix + 10n CGS matrix + 11n Bi-CGSTAB matrix + 10n QMR matrix + 16nc CHEBYSHEV matrix + 5n Table 2.2: Storage Requirements for the Methods in iteration i: n denotes the order of the matrix. cLess for implementations that do not recursively update the residual. 2. Gauss-Seidel Method • Typically faster convergence than Jacobi, but in general not competitive with the non- stationary methods. • Applicable to strictly diagonally dominant, or symmetric positive definite matrices. • Parallelization properties depend on structure of the coefficient matrix. Different order- ings of the unknowns have different degrees of parallelism; multi-color orderings may give almost full parallelism. • This is a special case of the SOR method, obtained by choosing ω = 1. 3. Successive Over-Relaxation (SOR) • Accelerates convergence of Gauss-Seidel (ω > 1, over-relaxation); may yield conver- gence when Gauss-Seidel fails (0 < ω < 1, under-relaxation). • Speed of convergence depends critically on ω; the optimal value for ω may be estimated from the spectral radius of the Jacobi iteration matrix under certain conditions. • Parallelization properties are the same as those of the Gauss-Seidel method. 4. Conjugate Gradient (CG) • Applicable to symmetric positive definite systems. • Speed of convergence depends on the condition number; if extremal eigenvalues are well-separated then superlinear convergence behavior can result. • Inner products act as synchronization points in a parallel environment. • Further parallel properties are largely independent of the coefficient matrix, but depend strongly on the structure the preconditioner. 5. Generalized Minimal Residual (GMRES) • Applicable to nonsymmetric matrices. • GMRES leads to the smallest residual for a fixed number of iteration steps, but these steps become increasingly expensive. • In order to limit the increasing storage requirments and work per iteration step, restart- ing is necessary. When to do so depends on A and the right-hand side; it requires skill and experience. • GMRES requires only matrix-vector products with the coefficient matrix.
- 39. 2.4. COMPUTATIONAL ASPECTS OF THE METHODS 29 • The number of inner products grows linearly with the iteration number, up to the restart point. In an implementation based on a simple Gram-Schmidt process the inner prod- ucts are independent, so together they imply only one synchronization point. A more stable implementation based on modified Gram-Schmidt orthogonalization has one synchronization point per inner product. 6. Biconjugate Gradient (BiCG) • Applicable to nonsymmetric matrices. • Requires matrix-vector products with the coefficient matrix and its transpose. This dis- qualifies the method for cases where the matrix is only given implicitly as an operator, since usually no corresponding transpose operator is available in such cases. • Parallelization properties are similar to those for CG; the two matrix vector products (as well as the preconditioning steps) are independent, so they can be done in parallel, or their communication stages can be packaged. 7. Quasi-Minimal Residual (QMR) • Applicable to nonsymmetric matrices. • Designed to avoid the irregular convergence behavior of BiCG, it avoids one of the two breakdown situations of BiCG. • If BiCG makes significant progress in one iteration step, then QMR delivers about the same result at the same step. But when BiCG temporarily stagnates or diverges, QMR may still further reduce the residual, albeit very slowly. • Computational costs per iteration are similar to BiCG, but slightly higher. The method requires the transpose matrix-vector product. • Parallelization properties are as for BiCG. 8. Conjugate Gradient Squared (CGS) • Applicable to nonsymmetric matrices. • Converges (diverges) typically about twice as fast as BiCG. • Convergence behavior is often quite irregular, which may lead to a loss of accuracy in the updated residual. Tends to diverge if the starting guess is close to the solution. • Computational costs per iteration are similar to BiCG, but the method doesn’t require the transpose matrix. • Unlike BiCG, the two matrix-vector products are not independent, so the number of synchronization points in a parallel environment is larger. 9. Biconjugate Gradient Stabilized (Bi-CGSTAB) • Applicable to nonsymmetric matrices. • Computational costs per iteration are similar to BiCG and CGS, but the method doesn’t require the transpose matrix. • An alternative for CGS that avoids the irregular convergence patterns of CGS while maintaining about the same speed of convergence; as a result we often observe less loss of accuracy in the updated residual. 10. Chebyshev Iteration • Applicable to nonsymmetric matrices (but presented in this book only for the symmetric case).
- 40. 30 CHAPTER 2. ITERATIVE METHODS • This method requires some explicit knowledge of the spectrum (or field of values); in the symmetric case the iteration parameters are easily obtained from the two extremal eigenvalues, which can be estimated either directly from the matrix, or from applying a few iterations of the Conjugate Gradient Method. • The computational structure is similar to that of CG, but there are no synchronization points. • The Adaptive Chebyshev method can be used in combination with methods as CG or GMRES, to continue the iteration once suitable bounds on the spectrum have been obtained from these methods. Selecting the “best” method for a given class of problems is largely a matter of trial and error. It also depends on how much storage one has available (GMRES), on the availability of AT (BiCG and QMR), and on how expensive the matrix vector products (and Solve steps with M) are in comparison to SAXPYs and inner products. If these matrix vector products are relatively expensive, and if sufficient storage is available then it may be attractive to use GMRES and delay restarting as much as possible. Table 2.1 shows the type of operations performed per iteration. Based on the particular prob- lem or data structure, the user may observe that a particular operation could be performed more efficiently. 2.5 A short history of Krylov methods1 Methods based on orthogonalization were developed by a number of authors in the early ’50s. Lanczos’ method [141] was based on two mutually orthogonal vector sequences, and his motivation came from eigenvalue problems. In that context, the most prominent feature of the method is that it reduces the original matrix to tridiagonal form. Lanczos later applied his method to solving linear systems, in particular symmetric ones [142]. An important property for proving convergence of the method when solving linear systems is that the iterates are related to the initial residual by multiplication with a polynomial in the coefficient matrix. The joint paper by Hestenes and Stiefel [121], after their independent discovery of the same method, is the classical description of the conjugate gradient method for solving linear systems. Although error-reduction properties are proved, and experiments showing premature convergence are reported, the conjugate gradient method is presented here as a direct method, rather than an iterative method. This Hestenes/Stiefel method is closely related to a reduction of the Lanczos method to sym- metric matrices, reducing the two mutually orthogonal sequences to one orthogonal sequence, but there is an important algorithmic difference. Whereas Lanczos used three-term recurrences, the method by Hestenes and Stiefel uses coupled two-term recurrences. By combining the two two- term recurrences (eliminating the “search directions”) the Lanczos method is obtained. A paper by Arnoldi [6] further discusses the Lanczos biorthogonalization method, but it also presents a new method, combining features of the Lanczos and Hestenes/Stiefel methods. Like the Lanczos method it is applied to nonsymmetric systems, and it does not use search directions. Like the Hestenes/Stiefel method, it generates only one, self-orthogonal sequence. This last fact, combined with the asymmetry of the coefficient matrix means that the method no longer effects a reduction to tridiagonal form, but instead one to upper Hessenberg form. Presented as “minimized iterations in the Galerkin method” this algorithm has become known as the Arnoldi algorithm. The conjugate gradient method received little attention as a practical method for some time, partly because of a misperceived importance of the finite termination property. Reid [178] pointed 1For a more detailed account of the early history of CG methods, we refer the reader to Golub and O’Leary [107] and Hestenes [122].
- 41. 2.6. SURVEY OF RECENT KRYLOV METHODS 31 out that the most important application area lay in sparse definite systems, and this renewed the interest in the method. Several methods have been developed in later years that employ, most often implicitly, the upper Hessenberg matrix of the Arnoldi method. For an overview and characterization of these orthogonal projection methods for nonsymmetric systems see Ashby, Manteuffel and Saylor [10], Saad and Schultz [187], and Jea and Young [124]. Fletcher [97] proposed an implementation of the Lanczos method, similar to the Conjugate Gra- dient method, with two coupled two-term recurrences, which he named the bi-conjugate gradient method (BiCG). 2.6 Survey of recent Krylov methods Research into the design of Krylov subspace methods for solving nonsymmetric linear systems is an active field of research and new methods are still emerging. In this book, we have included only the best known and most popular methods, and in particular those for which extensive computa- tional experience has been gathered. In this section, we shall briefly highlight some of the recent developments and other methods not treated here. A survey of methods up to about 1991 can be found in Freund, Golub and Nachtigal [105]. Two more recent reports by Meier-Yang [150] and Tong [195] have extensive numerical comparisons among various methods, including several more recent ones that have not been discussed in detail in this book. Several suggestions have been made to reduce the increase in memory and computational costs in GMRES. An obvious one is to restart (this one is included in §2.3.4): GMRES(m). Another approach is to restrict the GMRES search to a suitable subspace of some higher-dimensional Krylov subspace. Methods based on this idea can be viewed as preconditioned GMRES methods. The simplest ones exploit a fixed polynomial preconditioner (see Johnson, Micchelli and Paul [125], Saad [182], and Nachtigal, Reichel and Trefethen [158]). In more sophisticated approaches, the polynomial preconditioner is adapted to the iterations (Saad [186]), or the preconditioner may even be some other (iterative) method of choice (Van der Vorst and Vuik [207], Axelsson and Vassilevski [24]). Stagnation is prevented in the GMRESR method (Van der Vorst and Vuik [207]) by including LSQR steps in some phases of the process. In De Sturler and Fokkema [63], part of the optimality of GMRES is maintained in the hybrid method GCRO, in which the iterations of the preconditioning method are kept orthogonal to the iterations of the underlying GCR method. All these approaches have advantages for some problems, but it is far from clear a priori which strategy is preferable in any given case. Recent work has focused on endowing the BiCG method with several desirable properties: (1) avoiding breakdown; (2) avoiding use of the transpose; (3) efficient use of matrix-vector products; (4) smooth convergence; and (5) exploiting the work expended in forming the Krylov space with AT for further reduction of the residual. As discussed before, the BiCG method can have two kinds of breakdown: Lanczos breakdown (the underlying Lanczos process breaks down), and pivot breakdown (the tridiagonal matrix T implicitly generated in the underlying Lanczos process encounters a zero pivot when Gaussian elimination without pivoting is used to factor it). Although such exact breakdowns are very rare in practice, near breakdowns can cause severe numerical stability problems. The pivot breakdown is the easier one to overcome and there have been several approaches proposed in the literature. It should be noted that for symmetric matrices, Lanczos breakdown cannot occur and the only possible breakdown is pivot breakdown. The SYMMLQ and QMR methods discussed in this book circumvent pivot breakdown by solving least squares systems. Other methods tackling this problem can be found in Fletcher [97], Saad [180], Gutknecht [113], and Bank and Chan [29, 28]. Lanczos breakdown is much more difficult to eliminate. Recently, considerable attention has been given to analyzing the nature of the Lanczos breakdown (see Parlett [171], and
- 42. 32 CHAPTER 2. ITERATIVE METHODS Gutknecht [112, 115]), as well as various look-ahead techniques for remedying it (see Brezinski and Sadok [39], Brezinski, Zaglia and Sadok [40, 41], Freund and Nachtigal [101], Parlett [171], Nachtigal [159], Freund, Gutknecht and Nachtigal [100], Joubert [128], Freund, Golub and Nachtigal [105], and Gutknecht [112, 115]). However, the resulting algorithms are usually too complicated to give in template form (some codes of Freund and Nachtigal are available on netlib.) Moreover, it is still not possible to eliminate breakdowns that require look-ahead steps of arbitrary size (incurable breakdowns). So far, these methods have not yet received much practical use but some form of look-ahead may prove to be a crucial component in future methods. In the BiCG method, the need for matrix-vector multiplies with AT can be inconvenient as well as doubling the number of matrix-vector multiplies compared with CG for each increase in the degree of the underlying Krylov subspace. Several recent methods have been proposed to over- come this drawback. The most notable of these is the ingenious CGS method by Sonneveld [191] discussed earlier, which computes the square of the BiCG polynomial without requiring AT – thus obviating the need for AT . When BiCG converges, CGS is often an attractive, faster converging alternative. However, CGS also inherits (and often magnifies) the breakdown conditions and the irregular convergence of BiCG (see Van der Vorst [205]). CGS also generated interest in the possibility of product methods, which generate iterates corresponding to a product of the BiCG polynomial with another polynomial of the same de- gree, chosen to have certain desirable properties but computable without recourse to AT . The Bi-CGSTAB method of Van der Vorst [205] is such an example, in which the auxiliary polynomial is defined by a local minimization chosen to smooth the convergence behavior. Gutknecht [114] noted that Bi-CGSTAB could be viewed as a product of BiCG and GMRES(1), and he suggested combining BiCG with GMRES(2) for the even numbered iteration steps. This was anticipated to lead to better convergence for the case where the eigenvalues of A are complex. A more efficient and more robust variant of this approach has been suggested by Sleijpen and Fokkema in [189], where they describe how to easily combine BiCG with any GMRES(m), for modest m. Many other basic methods can also be squared. For example, by squaring the Lanczos pro- cedure, Chan, de Pillis and Van der Vorst [45] obtained transpose-free implementations of BiCG and QMR. By squaring the QMR method, Freund and Szeto [103] derived a transpose-free QMR squared method which is quite competitive with CGS but with much smoother convergence. Un- fortunately, these methods require an extra matrix-vector product per step (three instead of two) which makes them less efficient. In addition to Bi-CGSTAB, several recent product methods have been designed to smooth the convergence of CGS. One idea is to use the quasi-minimal residual (QMR) principle to obtain smoothed iterates from the Krylov subspace generated by other product methods. Freund [104] proposed such a QMR version of CGS, which he called TFQMR. Numerical experiments show that TFQMR in most cases retains the desirable convergence features of CGS while correcting its erratic behavior. The transpose free nature of TFQMR, its low computational cost and its smooth convergence behavior make it an attractive alternative to CGS. On the other hand, since the BiCG polynomial is still used, TFQMR breaks down whenever CGS does. One possible remedy would be to combine TFQMR with a look-ahead Lanczos technique but this appears to be quite complicated and no methods of this kind have yet appeared in the literature. Recently, Chan et. al. [46] derived a similar QMR version of Van der Vorst’s Bi-CGSTAB method, which is called QMRCGSTAB. These methods offer smoother convergence over CGS and Bi-CGSTAB with little additional cost. There is no clear best Krylov subspace method at this time, and there will never be a best overall Krylov subspace method. Each of the methods is a winner in a specific problem class, and the main problem is to identify these classes and to construct new methods for uncovered classes. The paper by Nachtigal, Reddy and Trefethen [157] shows that for any of a group of methods (CG, BiCG, GMRES, CGNE, and CGS), there is a class of problems for which a given method is the winner and another one is the loser. This shows clearly that there will be no ultimate method. The best we can hope for is some expert system that guides the user in his/her choice. Hence, iterative methods will never reach the robustness of direct methods, nor will they beat direct methods for
- 43. 2.6. SURVEY OF RECENT KRYLOV METHODS 33 all problems. For some problems, iterative schemes will be most attractive, and for others, direct methods (or multigrid). We hope to find suitable methods (and preconditioners) for classes of very large problems that we are yet unable to solve by any known method, because of CPU-restrictions, memory, convergence problems, ill-conditioning, et cetera.
- 44. 34 CHAPTER 2. ITERATIVE METHODS
- 45. Chapter 3 Preconditioners 3.1 The why and how The convergence rate of iterative methods depends on spectral properties of the coefficient matrix. Hence one may attempt to transform the linear system into one that is equivalent in the sense that it has the same solution, but that has more favorable spectral properties. A preconditioner is a matrix that effects such a transformation. For instance, if a matrix M approximates the coefficient matrix A in some way, the transformed system M−1 Ax = M−1 b has the same solution as the original system Ax = b, but the spectral properties of its coefficient matrix M−1 A may be more favorable. In devising a preconditioner, we are faced with a choice between finding a matrix M that approximates A, and for which solving a system is easier than solving one with A, or finding a matrix M that approximates A−1 , so that only multiplication by M is needed. The majority of preconditioners falls in the first category; a notable example of the second category will be discussed in §3.5. 3.1.1 Cost trade-off Since using a preconditioner in an iterative method incurs some extra cost, both initially for the setup, and per iteration for applying it, there is a trade-off between the cost of constructing and applying the preconditioner, and the gain in convergence speed. Certain preconditioners need little or no construction phase at all (for instance the SSOR preconditioner), but for others, such as incomplete factorizations, there can be substantial work involved. Although the work in scalar terms may be comparable to a single iteration, the construction of the preconditioner may not be vectorizable/parallelizable even if application of the preconditioner is. In that case, the initial cost has to be amortized over the iterations, or over repeated use of the same preconditioner in multiple linear systems. Most preconditioners take in their application an amount of work proportional to the number of variables. This implies that they multiply the work per iteration by a constant factor. On the other hand, the number of iterations as a function of the matrix size is usually only improved by a constant. Certain preconditioners are able to improve on this situation, most notably the modified incomplete factorizations and preconditioners based on multigrid techniques. On parallel machines there is a further trade-off between the efficacy of a preconditioner in the classical sense, and its parallel efficiency. Many of the traditional preconditioners have a large sequential component. 35
- 46. 36 CHAPTER 3. PRECONDITIONERS 3.1.2 Left and right preconditioning The above transformation of the linear system A → M−1 A is often not what is used in practice. For instance, the matrix M−1 A is not symmetric, so, even if A and M are, the conjugate gradients method is not immediately applicable to this system. The method as described in figure 2.5 reme- dies this by employing the M−1 -inner product for orthogonalization of the residuals. The theory of the cg method is then applicable again. All cg-type methods in this book, with the exception of QMR, have been derived with such a combination of preconditioned iteration matrix and correspondingly changed inner product. Another way of deriving the preconditioned conjugate gradients method would be to split the preconditioner as M = M1M2 and to transform the system as M−1 1 AM−1 2 (M2x) = M−1 1 b. If M is symmetric and M1 = Mt 2, it is obvious that we now have a method with a symmetric iteration matrix, hence the conjugate gradients method can be applied. Remarkably, the splitting of M is in practice not needed. By rewriting the steps of the method (see for instance Axelsson and Barker [14, pgs. 16,29] or Golub and Van Loan [108, §10.3]) it is usually possible to reintroduce a computational step solve u from Mu = v, that is, a step that applies the preconditioner in its entirety. There is a different approach to preconditioning, which is much easier to derive. Consider again the system. M−1 1 AM−1 2 (M2x) = M−1 1 b. The matrices M1 and M2 are called the left- and right preconditioners, respectively, and we can simpy apply an unpreconditioned iterative method to this system. Only two additional actions r0 ← M−1 1 r0 before the iterative process and xn ← M−1 2 xn after are necessary. Thus we arrive at the following schematic for deriving a left/right preconditioned iterative method from any of the symmetrically preconditioned methods in this book. 1. Take a preconditioned iterative method, and replace every occurence of M by I. 2. Remove any vectors from the algorithm that have become duplicates in the previous step. 3. Replace every occurrence of A in the method by M−1 1 AM−1 2 . 4. After the calculation of the initial residual, add the step r0 ← M−1 1 r0. 5. At the end of the method, add the step x ← M−1 2 x, where x is the final calculated solution. It should be noted that such methods cannot be made to reduce to the algorithms given in section 2.3 by such choices as M1 = Mt 2 or M1 = I.
- 47. 3.2. JACOBI PRECONDITIONING 37 3.2 Jacobi Preconditioning The simplest preconditioner consists of just the diagonal of the matrix: mi,j = ai,i if i = j 0 otherwise. This is known as the (point) Jacobi preconditioner. It is possible to use this preconditioner without using any extra storage beyond that of the matrix itself. However, division operations are usually quite costly, so in practice storage is allocated for the reciprocals of the matrix diagonal. This strategy applies to many preconditioners below. 3.2.1 Block Jacobi Methods Block versions of the Jacobi preconditioner can be derived by a partitioning of the variables. If the index set S = {1, . . . , n} is partitioned as S = i Si with the sets Si mutually disjoint, then mi,j = ai,j if i and j are in the same index subset 0 otherwise. The preconditioner is now a block-diagonal matrix. Often, natural choices for the partitioning suggest themselves: • In problems with multiple physical variables per node, blocks can be formed by grouping the equations per node. • In structured matrices, such as those from partial differential equations on regular grids, a partitioning can be based on the physical domain. Examples are a partitioning along lines in the 2D case, or planes in the 3D case. This will be discussed further in §3.4.3. • On parallel computers it is natural to let the partitioning coincide with the division of vari- ables over the processors. 3.2.2 Discussion Jacobi preconditioners need very little storage, even in the block case, and they are easy to imple- ment. Additionally, on parallel computers they don’t present any particular problems. On the other hand, more sophisticated preconditioners usually yield a larger improvement.1 3.3 SSOR preconditioning The SSOR preconditioner2 like the Jacobi preconditioner, can be derived from the coefficient ma- trix without any work. If the original, symmetric, matrix is decomposed as A = D + L + LT 1Under certain conditions, one can show that the point Jacobi algorithm is optimal, or close to optimal, in the sense of reducing the condition number, among all preconditioners of diagonal form. This was shown by Forsythe and Strauss for matrices with Property A [98], and by van der Sluis [196] for general sparse matrices. For extensions to block Jacobi preconditioners, see Demmel [65] and Elsner [94]. 2The SOR and Gauss-Seidel matrices are never used as preconditioners, for a rather technical reason. SOR- preconditioning with optimal ω maps the eigenvalues of the coefficient matrix to a circle in the complex plane; see Hageman and Young [119, §9.3]. In this case no polynomial acceleration is possible, i.e., the accelerating polynomial reduces to the trivial polynomial Pn(x) = xn, and the resulting method is simply the stationary SOR method. Recent research by Eier- mann and Varga [83] has shown that polynomial acceleration of SOR with suboptimal ω will yield no improvement over simple SOR with optimal ω.
- 48. 38 CHAPTER 3. PRECONDITIONERS in its diagonal, lower, and upper triangular part, the SSOR matrix is defined as M = (D + L)D−1 (D + L)T , or, parametrized by ω M(ω) = 1 2 − ω ( 1 ω D + L)( 1 ω D)−1 ( 1 ω D + L)T . The optimal value of the ω parameter, like the parameter in the SOR method, will reduce the number of iterations to a lower order. Specifically, for second order eliptic problems a spectral condition number κ(M−1 ωopt A) = O( κ(A)) is attainable (see Axelsson and Barker [14, §1.4]). In practice, however, the spectral information needed to calculate the optimal ω is prohibitively expensive to compute. The SSOR matrix is given in factored form, so this preconditioner shares many properties of other factorization-based methods (see below). For instance, its suitability for vector processors or parallel architectures depends strongly on the ordering of the variables. On the other hand, since this factorization is given a priori, there is no possibility of breakdown as in the construction phase of incomplete factorization methods. 3.4 Incomplete Factorization Preconditioners A broad class of preconditioners is based on incomplete factorizations of the coefficient matrix. We call a factorization incomplete if during the factorization process certain fill elements, nonzero elements in the factorization in positions where the original matrix had a zero, have been ignored. Such a preconditioner is then given in factored form M = LU with L lower and U upper triangular. The efficacy of the preconditioner depends on how well M−1 approximates A−1 . 3.4.1 Creating an incomplete factorization Incomplete factorizations are the first preconditioners we have encountered so far for which there is a non-trivial creation stage. Incomplete factorizations may break down (attempted division by zero pivot) or result in indefinite matrices (negative pivots) even if the full factorization of the same matrix is guaranteed to exist and yield a positive definite matrix. An incomplete factorization is guaranteed to exist for many factorization strategies if the orig- inal matrix is an M-matrix. This was originally proved by Meijerink and Van der Vorst [151]; see further Beauwens and Quenon [33], Manteuffel [146], and Van der Vorst [198]. In cases where pivots are zero or negative, strategies have been proposed such as substituting an arbitrary positive number (see Kershaw [131]), or restarting the factorization on A + αI for some positive value of α (see Manteuffel [146]). An important consideration for incomplete factorization preconditioners is the cost of the fac- torization process. Even if the incomplete factorization exists, the number of operations involved in creating it is at least as much as for solving a system with such a coefficient matrix, so the cost may equal that of one or more iterations of the iterative method. On parallel computers this problem is aggravated by the generally poor parallel efficiency of the factorization. Such factorization costs can be amortized if the iterative method takes many iterations, or if the same preconditioner will be used for several linear systems, for instance in successive time steps or Newton iterations. Solving a system with an incomplete factorization preconditioner Incomplete factorizations can be given in various forms. If M = LU (with L and U nonsingular triangular matrices), solving a system proceeds in the usual way (figure 3.1), but often incomplete
- 49. 3.4. INCOMPLETE FACTORIZATION PRECONDITIONERS 39 Let M = LU and y be given. for i = 1, 2, . . . zi = −1 ii (yi − j<i ijzj) for i = n, n − 1, n − 2, . . . xi = u−1 ii (zi − j>i uijxj) Figure 3.1: Preconditioner solve of a system Mx = y, with M = LU Let M = (D + L)(I + D−1 U) and y be given. for i = 1, 2, . . . zi = d−1 ii (yi − j<i ijzj) for i = n, n − 1, n − 2, . . . xi = zi − d−1 ii j>i uijxj Figure 3.2: Preconditioner solve of a system Mx = y, with M = (D + L)D−1 (D + U) = (D + L)(I + D−1 U). factorizations are given as M = (D +L)D−1 (D +U) (with D diagonal, and L and U now strictly triangular matrices, determined through the factorization process). In that case, one could use either of the following equivalent formulations for Mx = y: (D + L)z = y, (I + D−1 U)x = z or (I + LD−1 )z = y, (D + U)x = z. In either case, the diagonal elements are used twice (not three times as the formula for M would lead one to expect), and since only divisions with D are performed, storing D−1 explicitly is the practical thing to do. At the cost of some extra storage, one could store LD−1 or D−1 U, thereby saving some computation. Solving a system using the first formulation is outlined in figure 3.2. The second formulation is slightly harder to implement. 3.4.2 Point incomplete factorizations The most common type of incomplete factorization is based on taking a set S of matrix positions, and keeping all positions outside this set equal to zero during the factorization. The resulting factorization is incomplete in the sense that fill is supressed. The set S is usually chosen to encompass all positions (i, j) for which ai,j = 0. A position that is zero in A but not so in an exact factorization3 is called a fill position, and if it is outside S, the fill there is said to be “discarded”. Often, S is chosen to coincide with the set of nonzero positions in A, discarding all fill. This factorization type is called the ILU(0) factorization: the Incomplete LU factorization of level zero4 . 3To be precise, if we make an incomplete factorization M = LU, we refer to positions in L and U when we talk of positions in the factorization. The matrix M will have more nonzeros than L and U combined. 4The zero refers to the fact that only “level zero” fill is permitted, that is, nonzero elements of the original matrix. Fill levels are defined by calling an element of level k + 1 if it is caused by elements at least one of which is of level k. The first fill level is that caused by the original matrix elements.
- 50. 40 CHAPTER 3. PRECONDITIONERS Let S be the nonzero set {(i, j): aij = 0} for i = 1, 2, . . . set dii ← aii for i = 1, 2, . . . set dii ← 1/dii for j = i + 1, i + 2, . . . if (i, j) ∈ S and (j, i) ∈ S then set djj ← djj − ajidiiaij Figure 3.3: Construction of a D-ILU incomplete factorization preconditioner, storing the inverses of the pivots We can describe an incomplete factorization formally as for each k, i, j > k: ai,j ← ai,j − ai,ka−1 k,kak,j if (i, j) ∈ S ai,j otherwise. Meijerink and Van der Vorst [151] proved that, if A is an M-matrix, such a factorization exists for any choice of S, and gives a symmetric positive definite matrix if A is symmetric positive definite. Guidelines for allowing levels of fill were given by Meijerink and Van der Vorst in [152]. Fill-in strategies There are two major strategies for accepting or discarding fill-in, one structural, and one numerical. The structural strategy is that of accepting fill-in only to a certain level. As was already pointed out above, any zero location (i, j) in A filling in (say in step k) is assigned a fill level value fill(i, j) = 1 + max{fill(i, k), fill(k, i)}. (3.1) If aij was already nonzero, the level value is not changed. The numerical fill strategy is that of ‘drop tolerances’: fill is ignored if it is too small, for a suitable definition of ‘small’. Although this definition makes more sense mathematically, it is harder to implement in practice, since the amount of storage needed for the factorization is not easy to predict. See [20, 156] for discussions of preconditioners using drop tolerances. Simple cases: ILU(0) and D-ILU For the ILU(0) method, the incomplete factorization produces no nonzero elements beyond the sparsity structure of the original matrix, so that the preconditioner at worst takes exactly as much space to store as the original matrix. In a simplified version of ILU(0), called D-ILU (Pom- merell [173]), even less is needed. If not only we prohibit fill-in elements, but we also alter only the diagonal elements (that is, any alterations of off-diagonal elements are ignored5 ), we have the following situation. Splitting the coefficient matrix into its diagonal, lower triangular, and upper triangular parts as A = DA+LA+UA, the preconditioner can be written as M = (D+LA)D−1 (D+UA) where D is the diagonal matrix containing the pivots generated. Generating this preconditioner is described in figure 3.3. Since we use the upper and lower triangle of the matrix unchanged, only storage space for D is needed. In fact, in order to avoid division operations during the preconditioner solve stage we store D−1 rather than D. 5In graph theoretical terms, ILU(0) and D-ILU coincide if the matrix graph contains no triangles.
- 51. 3.4. INCOMPLETE FACTORIZATION PRECONDITIONERS 41 Remark: the resulting lower and upper factors of the preconditioner have only nonzero elements in the set S, but this fact is in general not true for the preconditioner M itself. The fact that the D-ILU preconditioner contains the off-diagonal parts of the original matrix was used by Eisenstat [90] to derive at a more efficient implementation of preconditioned CG. This new implementation merges the application of the tridiagonal factors of the matrix and the preconditioner, thereby saving a substantial number of operations per iteration. Special cases: central differences We will now consider the special case of a matrix derived from central differences on a Cartesian product grid. In this case the ILU(0) and D-ILU factorizations coincide, and, as remarked above, we only have to calculate the pivots of the factorization; other elements in the triangular factors are equal to off-diagonal elements of A. In the following we will assume a natural, line-by-line, ordering of the grid points. Letting i,j be coordinates in a regular 2D grid, it is easy to see that the pivot on grid point (i, j) is only determined by pivots on points (i − 1, j) and (i, j − 1). If there are n points on each of m grid lines, we get the following generating relations for the pivots: di,i = a1,1 if i = 1 ai,i − ai,i−1d−1 i−1ai−1,i if 1 < i ≤ n ai,i − ai,i−nd−1 i−nai−n,i if i = kn + 1 with k ≥ 1 ai,i − ai,i−1d−1 i−1ai−1,i − ai,i−nd−1 i−nai−n,i otherwise. Conversely, we can describe the factorization algorithmically as Initially: di,i = ai,i for all i for i = 1..nm do: di+1,i+1 = di+1,i+1 − ai+1,id−1 i,i ai,i+1 if there is no k such that i = kn di+n,i+n = di+n,i+n − ai+n,id−1 i,i ai,i+n if i + n ≤ nm In the above we have assumed that the variables in the problem are ordered according to the so-called “natural ordering”: a sequential numbering of the grid lines and the points within each grid line. Below we will encounter different orderings of the variables. Modified incomplete factorizations One modification to the basic idea of incomplete factorizations is as follows: If the product ai,ka−1 k,kak,j is nonzero, and fill is not allowed in position (i, j), instead of simply discarding this fill quantity subtract it from the diagonal element ai,i. Such a factorization scheme is usually called a “modified incomplete factorization”. Mathematically this corresponds to forcing the preconditioner to have the same rowsums as the original matrix. One reason for considering modified incomplete factorizations is the behavior of the spectral condition number of the preconditioned system. It was mentioned above that for second order elliptic equations the condition number of the coefficient matrix is O(h−2 ) as a func- tion of the discretization mesh width. This order of magnitude is preserved by simple incomplete factorizations, although usually a reduction by a large constant factor is obtained. Modified factorizations are of interest because, in combination with small perturbations, the spectral condition number of the preconditioned system can be of a lower order. It was first proved by Dupont, Kendall and Rachford [80] that a modified incomplete factorization of A + O(h2 )DA gives κ(M−1 A) = O(h−1 ) for the central difference case. More general proofs are given by Gustafsson [111], Axelsson and Barker [14, §7.2], and Beauwens [31, 32].
- 52. 42 CHAPTER 3. PRECONDITIONERS for k = 1, . . . , 2n − 1 do in parallel for i = max(1, k + 1 − n), min(n, k) j = k − i + 1 xi+(j−1)n ← Di+(j−1)nyi+(j−1)n −Li+(j−1)ni − 1 + (j − 1)nxi−1+(j−1)n −Li+(j−1)ni + (j − 2)nxi+(j−2)n Figure 3.4: Wavefront solution of (D + L)x = u from a central difference problem on a domain of n × n points. Instead of keeping row sums constant, one can also keep column sums constant. In compu- tational fluid mechanics this idea is justified with the argument that the material balance stays constant over all iterates. (Equivalently, one wishes to avoid ‘artificial diffusion’.) Appleyard and Cheshire [4] observed that if A and M have the same column sums, the choice x0 = M−1 b guarantees that the sum of the elements in r0 (the material balance error) is zero, and that all further ri have elements summing to zero. Modified incomplete factorizations can break down, especially when the variables are num- bered other than in the natural row-by-row ordering. This was noted by Chan and Kuo [50], and a full analysis was given by Eijkhout [85] and Notay [160]. A slight variant of modified incomplete factorizations consists of the class of “relaxed incom- plete factorizations”. Here the fill is multiplied by a parameter 0 < α < 1 before it is subtracted from the diagonal; see Ashcraft and Grimes [11], Axelsson and Lindskog [18, 19], Chan [44], Eijkhout [85], Notay [161], Stone [193], and Van der Vorst [202]. For the dangers of MILU in the presence of rounding error, see Van der Vorst [204]. Vectorization of the preconditioner solve At first it may appear that the sequential time of solving a factorization is of the order of the number of variables, but things are not quite that bad. Consider the special case of central differences on a regular domain of n × n points. The variables on any diagonal in the domain, that is, in locations (i, j) with i + j = k, depend only on those on the previous diagonal, that is, with i + j = k − 1. Therefore it is possible to process the operations on such a diagonal, or ‘wavefront’, in parallel (see figure 3.4), or have a vector computer pipeline them; see Van der Vorst [201, 203]. Another way of vectorizing the solution of the triangular factors is to use some form of expan- sion of the inverses of the factors. Consider for a moment a lower triangular matrix, normalized to the form I − L where L is strictly lower triangular). Its inverse can be given as either of the following two series: (I − L)−1 = I + L + L2 + L3 + · · · (I + L)(I + L2 )(I + L4 ) · · · (3.2) (The first series is called a “Neumann expansion”, the second an “Euler expansion”. Both series are finite, but their length prohibits practical use of this fact.) Parallel or vectorizable preconditioners can be derived from an incomplete factorization by taking a small number of terms in either series. Experiments indicate that a small number of terms, while giving high execution rates, yields almost the full precision of the more recursive triangular solution (see Axelsson and Eijkhout [15] and Van der Vorst [199]).
- 53. 3.4. INCOMPLETE FACTORIZATION PRECONDITIONERS 43 Let M = (I + L)D(I + U) and y be given. t ← y for k = 1, . . . , p t ← y − Lt x ← D−1 t, t ← x for k = 1, . . . , p x ← t − Ux Figure 3.5: Preconditioning step algorithm for a Neumann expansion M(p) ≈ M−1 of an incom- plete factorization M = (I + L)D(I + U). There are some practical considerations in implementing these expansion algorithms. For in- stance, because of the normalization the L in equation (3.2) is not LA. Rather, if we have a precon- ditioner (as described in section 3.4.2) described by A = DA + LA + UA, M = (D + LA)D−1 (D + UA), then we write M = (I + L)D(I + U), where L = LAD−1 , U = D−1 UA. Now we can choose whether or not to store the product LAD−1 . Doing so doubles the storage requirements for the matrix, not doing so means that separate multiplications by LA and D−1 have to be performed in the expansion. Suppose then that the products L = LAD−1 and U = D−1 UA have been stored. We then define M(p) ≈ M−1 by M(p) = p k=0 (−U)p D−1 p k=0 (−L)p , (3.3) and replace solving a system Mx = y for x by computing x = M(p) y. This algorithm is given in figure 3.5. Parallelizing the preconditioner solve The algorithms for vectorization outlined above can be used on parallel computers. For instance, variables on a wavefront can be processed in parallel, by dividing the wavefront over processors. More radical approaches for increasing the parallelism in incomplete factorizations are based on a renumbering of the problem variables. For instance, on rectangular domains one could start numbering the variables from all four corners simultaneously, thereby creating four simultaneous wavefronts, and therefore four-fold parallelism (see Dongarra, et al. [70], Van der Vorst [200, 202]). The most extreme case is the red/black ordering (or for more general matrices the multi-color ordering) which gives the absolute minimum number of sequential steps. Multi-coloring is also an attractive method for vector computers. Since points of one color are uncoupled, they can be processed as one vector; see Doi [67], Melhem [153], and Poole and Ortega [175]. However, for such ordering strategies there is usually a trade-off between the degree of paral- lelism and the resulting number of iterations. The reason for this is that a different ordering may give rise to a different error matrix, in particular the norm of the error matrix may vary considerably between orderings. See experimental results by Duff and Meurant [78] and a partial explanation of them by Eijkhout [84].
- 54. 44 CHAPTER 3. PRECONDITIONERS for i = 1, 2, . . . Xi ← Aii for i = 1, 2, . . . Let Yi ≈ X−1 i for j = i + 1, i + 2, . . . if Aij = 0 and Aji = 0, then Xj ← Xj − AjiYiAij Figure 3.6: Block version of a D-ILU factorization 3.4.3 Block factorization methods We can also consider block variants of preconditioners for accelerated methods. Block methods are normally feasible if the problem domain is a Cartesian product grid; in that case a natural division in lines (or planes in the 3-dimensional case), can be used for blocking, though incomplete factorizations are not as effective in the 3-dimensional case; see for instance Kettler [133]. In such a blocking scheme for Cartesian product grids, both the size and number of the blocks increases with the overall problem size. The idea behind block factorizations The starting point for an incomplete block factorization is a partitioning of the matrix, as mentioned in §3.2.1. Then an incomplete factorization is performed using the matrix blocks as basic entities (see Axelsson [12] and Concus, Golub and Meurant [56] as basic references). The most important difference with point methods arises in the inversion of the pivot blocks. Whereas inverting a scalar is easily done, in the block case two problems arise. First, inverting the pivot block is likely to be a costly operation. Second, initially the diagonal blocks of the matrix are likely to be be sparse and we would like to maintain this type of structure throughout the factorization. Hence the need for approximations of inverses arises. In addition to this, often fill-in in off-diagonal blocks is discarded altogether. Figure 3.6 describes an incomplete block factorization that is analogous to the D-ILU factorization (sec- tion 3.4.2) in that it only updates the diagonal blocks. As in the case of incomplete point factorizations, the existence of incomplete block methods is guaranteed if the coefficient matrix is an M-matrix. For a general proof, see Axelsson [13]. Approximate inverses In block factorizations a pivot block is generally forced to be sparse, typically of banded form, and that we need an approximation to its inverse that has a similar structure. Furthermore, this approximation should be easily computable, so we rule out the option of calculating the full inverse and taking a banded part of it. The simplest approximation to A−1 is the diagonal matrix D of the reciprocals of the diagonal of A: di,i = 1/ai,i. Other possibilities were considered by Axelsson and Eijkhout [15], Axelsson and Polman [21], Concus, Golub and Meurant [56], Eijkhout and Vassilevski [89], Kolotilina and Yeremin [140], and Meurant [154]. One particular example is given in figure 3.7. It has the attractive theoretical property that, if the original matrix is symmetric positive definite and a factorization with positive diagonal D can be made, the approximation to the inverse is again symmetric positive definite.
- 55. 3.4. INCOMPLETE FACTORIZATION PRECONDITIONERS 45 Let X be a banded matrix, factor X = (I + L)D−1 (I + U), let Y = (I − U)D(I − L) Figure 3.7: Algorithm for approximating the inverse of a banded matrix X1 = A1,1 for i ≥ 1 let Yi ≈ X−1 i Xi+1 = Ai+1,i+1 − Ai+1,iYiAi,i+1 Figure 3.8: Incomplete block factorization of a block tridiagonal matrix Banded approximations to the inverse of banded matrices have a theoretical justification. In the context of partial differential equations the diagonal blocks of the coefficient matrix are usually strongly diagonally dominant. For such matrices, the elements of the inverse have a size that is ex- ponentially decreasing in their distance from the main diagonal. See Demko, Moss and Smith [64] for a general proof, and Eijkhout and Polman [88] for a more detailed analysis in the M-matrix case. The special case of block tridiagonality In many applications, a block tridiagonal structure can be found in the coefficient matrix. Examples are problems on a 2D regular grid if the blocks correspond to lines of grid points, and problems on a regular 3D grid, if the blocks correspond to planes of grid points. Even if such a block tridiagonal structure does not arise naturally, it can be imposed by renumbering the variables in a Cuthill-McKee ordering [59]. Such a matrix has incomplete block factorizations of a particularly simple nature: since no fill can occur outside the diagonal blocks (Ai,i), all properties follow from our treatment of the pivot blocks. The generating recurrence for the pivot blocks also takes a simple form, see figure 3.8. After the factorization we are left with sequences of Xi block forming the pivots, and of Yi blocks approximating their inverses. Two types of incomplete block factorizations One reason that block methods are of interest is that they are potentially more suitable for vector computers and parallel architectures. Consider the block factorization A = (D + L)D−1 (D + U) = (D + L)(I + D−1 U) where D is the block diagonal matrix of pivot blocks,6 and L, U are the block lower and upper triangle of the factorization; they coincide with LA, UQ in the case of a block tridiagonal matrix. We can turn this into an incomplete factorization by replacing the block diagonal matrix of pivots D by the block diagonal matrix of incomplete factorization pivots X = diag(Xi), giving M = (X + L)(I + X−1 U) 6Writing (I + LAD−1)(D + UA) is equally valid, but in practice harder to implement.
- 56. 46 CHAPTER 3. PRECONDITIONERS For factorizations of this type (which covers all methods in Concus, Golub and Meurant [56] and Kolotilina and Yeremin [140]) solving a linear system means solving smaller systems with the Xi matrices. Alternatively, we can replace D by a the inverse of the block diagonal matrix of the approxi- mations to the inverses of the pivots, Y = diag(Yi), giving M = (Y −1 + L)(I + Y U). For this second type (which was discussed by Meurant [154], Axelsson and Polman [21] and Ax- elsson and Eijkhout [15]) solving a system with M entails multiplying by the Yi blocks. Therefore, the second type has a much higher potential for vectorizability. Unfortunately, such a factorization is theoretically more troublesome; see the above references or Eijkhout and Vassilevski [89]. 3.4.4 Blocking over systems of partial differential equations If the physical problem has several variables per grid point, that is, if there are several coupled partial differential equations, it is possible to introduce blocking in a natural way. Blocking of the equations (which gives a small number of very large blocks) was used by Axelsson and Gustafsson [17] for the equations of linear elasticity, and blocking of the variables per node (which gives many very small blocks) was used by Aarden and Karlsson [1] for the semiconductor equations. A systematic comparison of the two approaches was made by Bank, et al. [26]. 3.4.5 Incomplete LQ factorizations Saad [183] proposes to construct an incomplete LQ factorization of a general sparse matrix. The idea is to orthogonalize the rows of the matrix by a Gram-Schmidt process (note that in sparse matrices, most rows are typically orthogonal already, so that standard Gram-Schmidt may be not so bad as in general). Saad suggest dropping strategies for the fill-in produced in the orthogonalization process. It turns out that the resulting incomplete L factor can be viewed as the incomplete Choleski factor of the matrix AAT . Experiments show that using L in a CG process for the normal equations: L−1 AAT L−T y = b is effective for some relevant problems. 3.5 Polynomial preconditioners So far, we have described preconditioners in only one of two classes: those that approximate the coefficient matrix, and where linear systems with the preconditioner as coefficient matrix are easier to solve than the original system. Polynomial preconditioners can be considered as members of the second class of preconditioners: direct approximations of the inverse of the coefficient matrix. Suppose that the coefficient matrix A of the linear system is normalized to the form A = I −B, and that the spectral radius of B is less than one. Using the Neumann series, we can write the inverse of A as A−1 = ∞ k=0 Bk , so an approximation may be derived by truncating this infinite series. Since the iterative methods we are considering are already based on the idea of applying polynomials in the coefficient matrix to the initial residual, there are analytic connections between the basic method and polynomially accelerated one. Dubois, Greenbaum and Rodrigue [76] investigated the relationship between a basic method using a splitting A = M − N, and a polynomially preconditioned method with M−1 p = ( p−1 i=0 (I − M−1 A)i )M−1 . The basic result is that for classical methods, k steps of the polynomially preconditioned method are exactly equivalent to kp steps of the original method; for accelerated methods, specifically the
- 57. 3.6. OTHER PRECONDITIONERS 47 Chebyshev method, the preconditioned iteration can improve the number of iterations by at most a factor of p. Although there is no gain in the number of times the coefficient matrix is applied, polynomial preconditioning does eliminate a large fraction of the inner products and update operations, so there may be an overall increase in efficiency. Let us define a polynomial preconditioner more abstractly as any polynomial M = Pn(A) normalized to P(0) = 1. Now the choice of the best polynomial preconditioner becomes that of choosing the best polynomial that minimizes I − M−1 A . For the choice of the infinity norm we thus obtain Chebyshev polynomials, and they require estimates of both a lower and upper bound on the spectrum of A. These estimates may be derived from the conjugate gradient iteration itself; see §5.1. Since an accurate lower bound on the spectrum of A may be hard to obtain, Johnson, Micchelli and Paul [125] and Saad [182] propose least squares polynomials based on several weight functions. These functions only require an upper bound and this is easily computed, using for instance the “Gerschgorin bound” maxi j |Ai,j|; see [209, §1.4]. Experiments comparing Chebyshev and least squares polynomials can be found in Ashby, Manteuffel and Otto [8]. Application of polynomial preconditioning to symmetric indefinite problems is described by Ashby, Manteuffel and Saylor [9]. There the polynomial is chosen so that it transforms the system into a definite one. 3.6 Preconditioners from properties of the differential equation A number of preconditioners exist that derive their justification from properties of the underlying partial differential equation. We will cover some of them here (see also §5.5 and §5.4). These preconditioners usually involve more work than the types discussed above, however, they allow for specialized faster solution methods. 3.6.1 Preconditioning by the symmetric part In §2.3.4 we pointed out that conjugate gradient methods for non-selfadjoint systems require the storage of previously calculated vectors. Therefore it is somewhat remarkable that preconditioning by the symmetric part (A + AT )/2 of the coefficient matrix A leads to a method that does not need this extended storage. Such a method was proposed by Concus and Golub [55] and Widlund [214]. However, solving a system with the symmetric part of a matrix may be no easier than solving a system with the full matrix. This problem may be tackled by imposing a nested iterative method, where a preconditioner based on the symmetric part is used. Vassilevski [210] proved that the efficiency of this preconditioner for the symmetric part carries over to the outer method. 3.6.2 The use of fast solvers In many applications, the coefficient matrix is symmetric and positive definite. The reason for this is usually that the partial differential operator from which it is derived is self-adjoint, coercive, and bounded (see Axelsson and Barker [14, §3.2]). It follows that for the coefficient matrix A the following relation holds for any matrix B from a similar differential equation: c1 ≤ xT Ax xT Bx ≤ c2 for all x, where c1, c2 do not depend on the matrix size. The importance of this is that the use of B as a preconditioner gives an iterative method with a number of iterations that does not depend on the matrix size.
- 58. 48 CHAPTER 3. PRECONDITIONERS Thus we can precondition our original matrix by one derived from a different PDE, if one can be found that has attractive properties as preconditioner. The most common choice is to take a matrix from a separable PDE. A system involving such a matrix can be solved with various so-called “fast solvers”, such as FFT methods, cyclic reduction, or the generalized marching algorithm (see Dorr [74], Swarztrauber [194], Bank [25] and Bank and Rose [27]). As a simplest example, any elliptic operator can be preconditioned with the Poisson operator, giving the iterative method −∆(un+1 − un) = −(Lun − f). In Concus and Golub [58] a transformation of this method is considered to speed up the conver- gence. As another example, if the original matrix arises from −(a(x, y)ux)x − (b(x, y)uy)y = f, the preconditioner can be formed from −(˜a(x)ux)x − (˜b(y)uy)y = f. An extension to the non-self adjoint case is considered by Elman and Schultz [93]. Fast solvers are attractive in that the number of operations they require is (slightly higher than) of the order of the number of variables. Coupled with the fact that the number of iterations in the resulting preconditioned iterative methods is independent of the matrix size, such methods are close to optimal. However, fast solvers are usually only applicable if the physical domain is a rectangle or other Cartesian product structure. (For a domain consisting of a number of such pieces, domain decomposition methods can be used; see §5.4). 3.6.3 Alternating Direction Implicit methods The Poisson differential operator can be split in a natural way as the sum of two operators: L = L1 + L2, where L1 = − ∂2 ∂x2 , L2 = − ∂2 ∂y2 . Now let L1, L2 be discretized representations of L1, L2. Based on the observation that L1 + L2 = (I + L1)(I + L2) − I − L1L2, iterative schemes such as (1 + αL1)(1 + αL2)u(m+1) = [(1 + βL1)(1 + βL2)] u(m) with suitable choices of α and β have been proposed. This alternating direction implicit, or ADI, method was first proposed as a solution method for parabolic equations. The u(m) are then approximations on subsequent time steps. However, it can also be used for the steady state, that is, for solving elliptic equations. In that case, the u(m) become subsequent iterates; see D’Yakonov [81], Fairweather, Gourlay and Mitchell [96], Hadjidimos [118], and Peaceman and Rachford [172]. Generalization of this scheme to variable coefficients or fourth order elliptic problems is relatively straightforward. The above method is implicit since it requires systems solutions, and it alternates the x and y (and if necessary z) directions. It is attractive from a practical point of view (although mostly on tensor product grids), since solving a system with, for instance, a matrix I + αL1 entails only a number of uncoupled tridiagonal solutions. These need very little storage over that needed for the matrix, and they can be executed in parallel, or one can vectorize over them. A theoretical reason that ADI preconditioners are of interest is that they can be shown to be spectrally equivalent to the original coefficient matrix. Hence the number of iterations is bounded independent of the condition number. However, there is a problem of data distribution. For vector computers, either the system solu- tion with L1 or with L2 will involve very large strides: if columns of variables in the grid are stored
- 59. 3.6. OTHER PRECONDITIONERS 49 contiguously, only the solution with L1 will involve contiguous data. For the L2 the stride equals the number of variables in a column. On parallel machines an efficient solution is possible if the processors are arranged in a Px ×Py grid. During, e.g., the L1 solve, every processor row then works independently of other rows. Inside each row, the processors can work together, for instance using a Schur complement method. With sufficient network bandwidth this will essentially reduce the time to that for solving any of the subdomain systems plus the time for the interface system. Thus, this method will be close to optimal.
- 60. 50 CHAPTER 3. PRECONDITIONERS
- 61. Chapter 4 Related Issues 4.1 Complex Systems Conjugate gradient methods for real symmetric systems can be applied to complex Hermitian sys- tems in a straightforward manner. For non-Hermitian complex systems we distinguish two cases. In general, for any coefficient matrix a CGNE method is possible, that is, a conjugate gradients method on the normal equations AH Ax = AH b, or one can split the system into real and complex parts and use a method such as GMRES on the resulting real nonsymmetric system. However, in certain practical situations the complex system is non-Hermitian but symmetric. Complex symmetric systems can be solved by a classical conjugate gradient or Lanczos method, that is, with short recurrences, if the complex inner product (x, y) = ¯xT y is replaced by (x, y) = xT y. Like the BiConjugate Gradient method, this method is susceptible to breakdown, that is, it can happen that xT x = 0 for x = 0. A look-ahead strategy can remedy this in most cases (see Freund [99] and Van der Vorst and Melissen [206]). 4.2 Stopping Criteria An iterative method produces a sequence {x(i) } of vectors converging to the vector x satisfying the n × n system Ax = b. To be effective, a method must decide when to stop. A good stopping criterion should 1. identify when the error e(i) ≡ x(i) − x is small enough to stop, 2. stop if the error is no longer decreasing or decreasing too slowly, and 3. limit the maximum amount of time spent iterating. For the user wishing to read as little as possible, the following simple stopping criterion will likely be adequate. The user must supply the quantities maxit, b , stop tol, and preferably also A : • The integer maxit is the maximum number of iterations the algorithm will be permitted to perform. • The real number A is a norm of A. Any reasonable (order of magnitude) approximation of the absolute value of the largest entry of the matrix A will do. • The real number b is a norm of b. Again, any reasonable approximation of the absolute value of the largest entry of b will do. 51
- 62. 52 CHAPTER 4. RELATED ISSUES • The real number stop tol measures how small the user wants the residual r(i) = Ax(i) − b of the ultimate solution x(i) to be. One way to choose stop tol is as the approximate uncertainty in the entries of A and b relative to A and b , respectively. For example, choosing stop tol 10−6 means that the user considers the entries of A and b to have errors in the range ±10−6 A and ±10−6 b , respectively. The algorithm will compute x no more accurately than its inherent uncertainty warrants. The user should choose stop tol less than one and greater than the machine precision ε.1 Here is the algorithm: i = 0 repeat i = i + 1 Compute the approximate solution x(i) . Compute the residual r(i) = Ax(i) − b. Compute r(i) and x(i) . until i ≥ maxit or r(i) ≤ stop tol · ( A · x(i) + b ). Note that if x(i) does not change much from step to step, which occurs near convergence, then x(i) need not be recomputed. If A is not available, the stopping criterion may be replaced with the generally stricter criterion until i ≥ maxit or r(i) ≤ stop tol · b . In either case, the final error bound is e(i) ≤ A−1 · r(i) . If an estimate of A−1 is available, one may also use the stopping criterion until i ≥ maxit or r(i) ≤ stop tol · x(i) / A−1 , which guarantees that the relative error e(i) / x(i) in the computed solution is bounded by stop tol. 4.2.1 More Details about Stopping Criteria Ideally we would like to stop when the magnitudes of entries of the error e(i) = x(i) −x fall below a user-supplied threshold. But e(i) is hard to estimate directly, so we use the residual r(i) = Ax(i) −b instead, which is more readily computed. The rest of this section describes how to measure the sizes of vectors e(i) and r(i) , and how to bound e(i) in terms of r(i) . We will measure errors using vector and matrix norms. The most common vector norms are: x ∞ ≡ maxj |xj| , x 1 ≡ j |xj| , and x 2 ≡ ( j |xj|2 )1/2 . For some algorithms we may also use the norm x B,α ≡ Bx α, where B is a fixed nonsingular matrix and α is one of ∞, 1, or 2. Corresponding to these vector norms are three matrix norms: A ∞ ≡ maxj k |aj,k| , A 1 ≡ maxk j |aj,k| , and A F ≡ ( jk |aj,k|2 )1/2 , as well as A B,α ≡ BAB−1 α. We may also use the matrix norm A 2 = (λmax(AAT ))1/2 , where λmax denotes the largest eigenvalue. Henceforth x and A will refer to any mutually 1 On a machine with IEEE Standard Floating Point Arithmetic, ε = 2−24 ≈ 10−7 in single precision, and ε = 2−53 ≈ 10−16 in double precision.
- 63. 4.2. STOPPING CRITERIA 53 consistent pair of the above. ( x 2 and A F , as well as x 2 and A 2, both form mutually consistent pairs.) All these norms satisfy the triangle inequality x + y ≤ x + y and A + B ≤ A + B , as well as Ax ≤ A · x for mutually consistent pairs. (For more details on the properties of norms, see Golub and Van Loan [108].) One difference between these norms is their dependence on dimension. A vector x of length n with entries uniformly distributed between 0 and 1 will satisfy x ∞ ≤ 1, but x 2 will grow like√ n and x 1 will grow like n. Therefore a stopping criterion based on x 1 (or x 2) may have to be permitted to grow proportional to n (or √ n) in order that it does not become much harder to satisfy for large n. There are two approaches to bounding the inaccuracy of the computed solution to Ax = b. Since e(i) , which we will call the forward error, is hard to estimate directly, we introduce the backward error, which allows us to bound the forward error. The normwise backward error is de- fined as the smallest possible value of max{ δA / A , δb / b } where x(i) is the exact solution of (A + δA)x(i) = (b + δb) (here δA denotes a general matrix, not δ times A; the same goes for δb). The backward error may be easily computed from the residual r(i) = Ax(i) −b; we show how below. Provided one has some bound on the inverse of A, one can bound the forward error in terms of the backward error via the simple equality e(i) = x(i) − x = A−1 (Ax(i) − b) = A−1 r(i) , which implies e(i) ≤ A−1 · r(i) . Therefore, a stopping criterion of the form “stop when r(i) ≤ τ” also yields an upper bound on the forward error e(i) ≤ τ · A−1 . (Sometimes we may prefer to use the stricter but harder to estimate bound e(i) ≤ |A−1 | · |r(i) | ; see §4.2.3. Here |X| is the matrix or vector of absolute values of components of X.) The backward error also has a direct interpretation as a stopping criterion, in addition to supplying a bound on the forward error. Recall that the backward error is the smallest change max{ δA / A , δb / b } to the problem Ax = b that makes x(i) an exact solution of (A + δA)x(i) = b + δb. If the original data A and b have errors from previous computations or mea- surements, then it is usually not worth iterating until δA and δb are even smaller than these errors. For example, if the machine precision is ε, it is not worth making δA ≤ ε A and δb ≤ ε b , because just rounding the entries of A and b to fit in the machine creates errors this large. Based on this discussion, we will now consider some stopping criteria and their properties. Above we already mentioned Criterion 1. r(i) ≤ S1 ≡ stop tol · ( A · x(i) + b ). This is equivalent to asking that the backward error δA and δb described above satisfy δA ≤ stop tol · A and δb ≤ stop tol · b . This criterion yields the forward error bound e(i) ≤ A−1 · r(i) ≤ stop tol · A−1 · ( A · x(i) + b ) . The second stopping criterion we discussed, which does not require A , may be much more stringent than Criterion 1: Criterion 2. r(i) ≤ S2 ≡ stop tol · b . This is equivalent to asking that the backward error δA and δb satisfy δA = 0 and δb ≤ tol · b . One difficulty with this method is that if A · x b , which can only occur if A is very ill-conditioned and x nearly lies in the null space of A, then it may be difficult for any method to satisfy the stopping criterion. To see that A must be very ill-conditioned, note that 1 A · x b = A · A−1 b b ≤ A · A−1 . This criterion yields the forward error bound e(i) ≤ A−1 · r(i) ≤ stop tol · A−1 · b
- 64. 54 CHAPTER 4. RELATED ISSUES If an estimate of A−1 is available, one can also just stop when the upper bound on the error A−1 · r(i) falls below a threshold. This yields the third stopping criterion: Criterion 3. r(i) ≤ S3 ≡ stop tol · x(i) / A−1 . This stopping criterion guarantees that e(i) x(i) ≤ A−1 · r(i) x(i) ≤ stop tol , permitting the user to specify the desired relative accuracy stop tol in the computed solution x(i) . One drawback to Criteria 1 and 2 is that they usually treat backward errors in each component of δA and δb equally, since most norms δA and δb measure each entry of δA and δb equally. For example, if A is sparse and δA is dense, this loss of possibly important structure will not be reflected in δA . In contrast, the following stopping criterion gives one the option of scaling each component δaj,k and δbj differently, including the possibility of insisting that some entries be zero. The cost is an extra matrix-vector multiply: Criterion 4. S4 ≡ maxj(|r(i) |j/(E · |x(i) | + f)j) ≤ stop tol. Here E is a user-defined matrix of nonnegative entries, f is a user-defined vector of nonnegative entries, and |z| denotes the vector of absolute values of the entries of z. If this criterion is satisfied, it means there are a δA and a δb such that (A + δA)x(i) = b + δb, with |δaj,k| ≤ tol · ej,k, and |δbj| ≤ tol · fj for all j and k. By choosing E and f, the user can vary the way the backward error is measured in the stopping criterion. For example, choosing ej,k = A ∞ and fj = b ∞ makes the stopping criterion r(i) ∞/(n A ∞ x(i) ∞ + b ∞), which is essentially the same as Criterion 1. Choosing ej,k = |aj,k| and fj = |bj| makes the stopping criterion measure the componentwise relative backward error, i.e., the smallest relative perturbations in any component of A and b which is necessary to make x(i) an exact solution. This tighter stopping criterion requires, among other things, that δA have the same sparsity pattern as A. Other choices of E and f can be used to reflect other structured uncertainties in A and b. This criterion yields the forward error bound e(i) ∞ ≤ |A−1 | · |r(i) | ≤ S4 · |A−1 |(E|x(i) | + f) ∞ where |A−1 | is the matrix of absolute values of entries of A−1 . Finally, we mention one more criterion, not because we recommend it, but because it is widely used. We mention it in order to explain its potential drawbacks: Dubious Criterion 5. r(i) ≤ S5 ≡ stop tol · r(0) . This commonly used criterion has the disadvantage of depending too strongly on the initial solution x(0) . If x(0) = 0, a common choice, then r(0) = b. Then this criterion is equivalent to Criterion 2 above, which may be difficult to satisfy for any algorithm if b A · x . On the other hand, if x(0) is very large and very inaccurate, then r(0) will be very large and S5 will be artificially large; this means the iteration may stop too soon. This criterion yields the forward error bound e(i) ≤ S5 · A−1 .
- 65. 4.2. STOPPING CRITERIA 55 4.2.2 When r(i) or r(i) is not readily available It is possible to design an iterative algorithm for which r(i) = Ax(i) − b or r(i) is not directly available, although this is not the case for any algorithms in this book. For completeness, however, we discuss stopping criteria in this case. For example, if ones “splits” A = M − N to get the iterative method x(i) = M−1 Nx(i−1) + M−1 b ≡ Gx(i−1) + c, then the natural residual to compute is ˆr(i) = x(i) − Gx(i) − c = M−1 (Ax(i) − b) = M−1 r(i) . In other words, the residual ˆr(i) is the same as the residual of the preconditioned system M−1 Ax = M−1 b. In this case, it is hard to interpret ˆr(i) as a backward error for the original system Ax = b, so we may instead derive a forward error bound e(i) = A−1 M ˆr(i) ≤ A−1 M · ˆr(i) . Using this as a stopping criterion requires an estimate of A−1 M . In the case of methods based on splitting A = M − N, we have A−1 M = (M − N)−1 M = (I − G)−1 , and A−1 M = (I − G)−1 ≤ 1/(1 − G ). Another example is an implementation of the preconditioned conjugate gradient algorithm which computes r(i) M−1/2,2 = (r(i)T M−1 r(i) )1/2 instead of r(i) 2 (the implementation in this book computes the latter). Such an implementation could use the stopping criterion r(i) M−1/2,2/ r(0) M−1/2,2 ≤ tol as in Criterion 5. We may also use it to get the forward error bound e(i) ≤ A−1 M1/2 · r(i) M−1/2,2, which could also be used in a stopping criterion. 4.2.3 Estimating A−1 Bounds on the error e(i) inevitably rely on bounds for A−1 , since e(i) = A−1 r(i) . There is a large number of problem dependent ways to estimate A−1 ; we mention a few here. When a splitting A = M − N is used to get an iteration x(i) = M−1 Nx(i−1) + M−1 b = Gx(i−1) + c, then the matrix whose inverse norm we need is I − G. Often, we know how to estimate G if the splitting is a standard one such as Jacobi or SOR, and the matrix A has special characteristics such as Property A. Then we may estimate (I − G)−1 ≤ 1/(1 − G ). When A is symmetric positive definite, and Chebyshev acceleration with adaptation of param- eters is being used, then at each step the algorithm estimates the largest and smallest eigenvalues λmax(A) and λmin(A) of A anyway. Since A is symmetric positive definite, A−1 2 = λ−1 min(A). This adaptive estimation is often done using the Lanczos algorithm (see section 5.1), which can usually provide good estimates of the largest (rightmost) and smallest (leftmost) eigenvalues of a symmetric matrix at the cost of a few matrix-vector multiplies. For general nonsymmetric A, we may apply the Lanczos method to AAT or AT A, and use the fact that A−1 2 = 1/λ 1/2 min(AAT ) = 1/λ 1/2 min(AT A). It is also possible to estimate A−1 ∞ provided one is willing to solve a few systems of linear equations with A and AT as coefficient matrices. This is often done with dense linear system solvers, because the extra cost of these systems is O(n2 ), which is small compared to the cost O(n3 ) of the LU decomposition (see Hager [120], Higham [123] and Anderson, et al. [3]). This is not the case for iterative solvers, where the cost of these solves may well be several times as much as the original linear system. Still, if many linear systems with the same coefficient matrix and differing right-hand-sides are to be solved, it is a viable method. The approach in the last paragraph also lets us estimate the alternate error bound e(i) ∞ ≤ |A−1 | · |r(i) | ∞. This may be much smaller than the simpler A−1 ∞ · r(i) ∞ in the case where the rows of A are badly scaled; consider the case of a diagonal matrix A with widely varying diagonal entries. To compute |A−1 | · |r(i) | ∞, let R denote the diagonal matrix with diagonal entries equal to the entries of |r(i) |; then |A−1 | · |r(i) | ∞ = A−1 R ∞ (see Arioli, Demmel and Duff [5]). A−1 R ∞ can be estimated using the technique in the last paragraph since multiplying by A−1 R or (A−1 R)T = RT A−T is no harder than multiplying by A−1 and A−T and also by R, a diagonal matrix.
- 66. 56 CHAPTER 4. RELATED ISSUES 4.2.4 Stopping when progress is no longer being made In addition to limiting the total amount of work by limiting the maximum number of iterations one is willing to do, it is also natural to consider stopping when no apparent progress is being made. Some methods, such as Jacobi and SOR, often exhibit nearly monotone linear convergence, at least after some initial transients, so it is easy to recognize when convergence degrades. Other methods, like the conjugate gradient method, exhibit “plateaus” in their convergence, with the residual norm stagnating at a constant value for many iterations before decreasing again; in principle there can be many such plateaus (see Greenbaum and Strakos [109]) depending on the problem. Still other methods, such as CGS, can appear wildly nonconvergent for a large number of steps before the residual begins to decrease; convergence may continue to be erratic from step to step. In other words, while it is a good idea to have a criterion that stops when progress towards a solution is no longer being made, the form of such a criterion is both method and problem depen- dent. 4.2.5 Accounting for floating point errors The error bounds discussed in this section are subject to floating point errors, most of which are innocuous, but which deserve some discussion. The infinity norm x ∞ = maxj |xj| requires the fewest floating point operations to compute, and cannot overflow or cause other exceptions if the xj are themselves finite2 . On the other hand, computing x 2 = ( j |xj|2 )1/2 in the most straightforward manner can easily overflow or lose accuracy to underflow even when the true result is far from either the overflow or underflow thresh- olds. For this reason, a careful implementation for computing x 2 without this danger is available (subroutine snrm2 in the BLAS [71] [143]), but it is more expensive than computing x ∞. Now consider computing the residual r(i) = Ax(i) − b by forming the matrix-vector product Ax(i) and then subtracting b, all in floating point arithmetic with relative precision ε. A standard error analysis shows that the error δr(i) in the computed r(i) is bounded by δr(i) ≤ O(ε)( A · x(i) + b ), where O(ε) is typically bounded by nε, and usually closer to √ nε. This is why one should not choose stop tol ≤ ε in Criterion 1, and why Criterion 2 may not be satisfied by any method. This uncertainty in the value of r(i) induces an uncertainty in the error e(i) = A−1 r(i) of at most O(ε) A−1 · ( A · x(i) + b ). A more refined bound is that the error (δr(i) )j in the jth component of r(i) is bounded by O(ε) times the jth component of |A| · |x(i) | + |b|, or more tersely |δr(i) | ≤ O(ε)(|A| · |x(i) | + |b|). This means the uncertainty in e(i) is really bounded by O(ε) |A−1 |·(|A|·|x(i) |+|b|) . This last quantity can be estimated inexpensively provided solving systems with A and AT as coefficient matrices is inexpensive (see the last paragraph of §4.2.3). Both these bounds can be severe overestimates of the uncertainty in e(i) , but examples exist where they are attainable. 4.3 Data Structures The efficiency of any of the iterative methods considered in previous sections is determined primar- ily by the performance of the matrix-vector product and the preconditioner solve, and therefore on the storage scheme used for the matrix and the preconditioner. Since iterative methods are typically used on sparse matrices, we will review here a number of sparse storage formats. Often, the storage scheme used arises naturally from the specific application problem. In this section we will review some of the more popular sparse matrix formats that are used in numerical software packages such as ITPACK [139] and NSPCG [164]. After surveying the various formats, we demonstrate how the matrix-vector product and an incomplete factorization solve are formulated using two of the sparse matrix formats. 2IEEE standard floating point arithmetic permits computations with ±∞ and NaN, or Not-a-Number, symbols.
- 67. 4.3. DATA STRUCTURES 57 4.3.1 Survey of Sparse Matrix Storage Formats If the coefficient matrix A is sparse, large-scale linear systems of the form Ax = b can be most efficiently solved if the zero elements of A are not stored. Sparse storage schemes allocate con- tiguous storage in memory for the nonzero elements of the matrix, and perhaps a limited number of zeros. This, of course, requires a scheme for knowing where the elements fit into the full matrix. There are many methods for storing the data (see for instance Saad [185] and Eijkhout [86]). Here we will discuss Compressed Row and Column Storage, Block Compressed Row Storage, Diagonal Storage, Jagged Diagonal Storage, and Skyline Storage. Compressed Row Storage (CRS) The Compressed Row and Column (in the next section) Storage formats are the most general: they make absolutely no assumptions about the sparsity structure of the matrix, and they don’t store any unnecessary elements. On the other hand, they are not very efficient, needing an indirect addressing step for every single scalar operation in a matrix-vector product or preconditioner solve. The Compressed Row Storage (CRS) format puts the subsequent nonzeros of the matrix rows in contiguous memory locations. Assuming we have a nonsymmetric sparse matrix A, we cre- ate 3 vectors: one for floating-point numbers (val), and the other two for integers (col ind, row ptr). The val vector stores the values of the nonzero elements of the matrix A, as they are traversed in a row-wise fashion. The col ind vector stores the column indexes of the el- ements in the val vector. That is, if val(k) = ai,j then col ind(k) = j. The row ptr vector stores the locations in the val vector that start a row, that is, if val(k) = ai,j then row ptr(i) ≤ k < row ptr(i + 1). By convention, we define row ptr(n + 1) = nnz + 1, where nnz is the number of nonzeros in the matrix A. The storage savings for this approach is significant. Instead of storing n2 elements, we need only 2nnz + n + 1 storage locations. As an example, consider the nonsymmetric matrix A defined by A = 10 0 0 0 −2 0 3 9 0 0 0 3 0 7 8 7 0 0 3 0 8 7 5 0 0 8 0 9 9 13 0 4 0 0 2 −1 . (4.1) The CRS format for this matrix is then specified by the arrays {val, col ind, row ptr} given below val 10 -2 3 9 3 7 8 7 3 · · · 9 13 4 2 -1 col ind 1 5 1 2 6 2 3 4 1 · · · 5 6 2 5 6 row ptr 1 3 6 9 13 17 20 . If the matrix A is symmetric, we need only store the upper (or lower) triangular portion of the matrix. The trade-off is a more complicated algorithm with a somewhat different pattern of data access. Compressed Column Storage (CCS) Analogous to Compressed Row Storage there is Compressed Column Storage (CCS), which is also called the Harwell-Boeing sparse matrix format [77]. The CCS format is identical to the CRS format except that the columns of A are stored (traversed) instead of the rows. In other words, the CCS format is the CRS format for AT . The CCS format is specified by the 3 arrays {val, row ind, col ptr}, where row ind stores the row indices of each nonzero, and col ptr stores the index of the elements in val which start a column of A. The CCS format for the matrix A in (4.1) is given by
- 68. 58 CHAPTER 4. RELATED ISSUES val 10 3 3 9 7 8 4 8 8 · · · 9 2 3 13 -1 row ind 1 2 4 2 3 5 6 3 4 · · · 5 6 2 5 6 col ptr 1 4 8 10 13 17 20 . Block Compressed Row Storage (BCRS) If the sparse matrix A is comprised of square dense blocks of nonzeros in some regular pattern, we can modify the CRS (or CCS) format to exploit such block patterns. Block matrices typically arise from the discretization of partial differential equations in which there are several degrees of freedom associated with a grid point. We then partition the matrix in small blocks with a size equal to the number of degrees of freedom, and treat each block as a dense matrix, even though it may have some zeros. If nb is the dimension of each block and nnzb is the number of nonzero blocks in the n × n matrix A, then the total storage needed is nnz = nnzb × n2 b. The block dimension nd of A is then defined by nd = n/nb. Similar to the CRS format, we require 3 arrays for the BCRS format: a rectangular array for floating-point numbers ( val(1 : nnzb,1 : nb,1 : nb)) which stores the nonzero blocks in (block) row-wise fashion, an integer array (col ind(1 : nnzb)) which stores the actual column indices in the original matrix A of the (1, 1) elements of the nonzero blocks, and a pointer array (row blk(1 : nd + 1)) whose entries point to the beginning of each block row in val(:,:,:) and col ind(:). The savings in storage locations and reduction in indirect addressing for BCRS over CRS can be significant for matrices with a large nb. Compressed Diagonal Storage (CDS) If the matrix A is banded with bandwidth that is fairly constant from row to row, then it is worth- while to take advantage of this structure in the storage scheme by storing subdiagonals of the matrix in consecutive locations. Not only can we eliminate the vector identifying the column and row, we can pack the nonzero elements in such a way as to make the matrix-vector product more efficient. This storage scheme is particularly useful if the matrix arises from a finite element or finite difference discretization on a tensor product grid. We say that the matrix A = (ai,j) is banded if there are nonnegative constants p, q, called the left and right halfbandwidth, such that ai,j = 0 only if i − p ≤ j ≤ i + q. In this case, we can allocate for the matrix A an array val(1:n,-p:q). The declaration with reversed dimensions (-p:q,n) corresponds to the LINPACK band format [72], which unlike CDS, does not allow for an efficiently vectorizable matrix-vector multiplication if p + q is small. Usually, band formats involve storing some zeros. The CDS format may even contain some ar- ray elements that do not correspond to matrix elements at all. Consider the nonsymmetric matrix A defined by A = 10 −3 0 0 0 0 3 9 6 0 0 0 0 7 8 7 0 0 0 0 8 7 5 0 0 0 0 9 9 13 0 0 0 0 2 −1 . (4.2) Using the CDS format, we store this matrix A in an array of dimension (6,-1:1) using the mapping val(i, j) = ai,i+j. (4.3) Hence, the rows of the val(:,:) array are
- 69. 4.3. DATA STRUCTURES 59 val(:,-1) 0 3 7 8 9 2 val(:, 0) 10 9 8 7 9 -1 val(:,+1) -3 6 7 5 13 0 . Notice the two zeros corresponding to non-existing matrix elements. A generalization of the CDS format more suitable for manipulating general sparse matrices on vector supercomputers is discussed by Melhem in [153]. This variant of CDS uses a stripe data structure to store the matrix A. This structure is more efficient in storage in the case of varying bandwidth, but it makes the matrix-vector product slightly more expensive, as it involves a gather operation. As defined in [153], a stripe in the n × n matrix A is a set of positions S = {(i, σ(i)); i ∈ I ⊆ In}, where In = {1, . . . , n} and σ is a strictly increasing function. Specifically, if (i, σ(i)) and (j, σ(j)) are in S, then i < j → σ(i) < σ(j). When computing the matrix-vector product y = Ax using stripes, each (i, σk(i)) element of A in stripe Sk is multiplied with both xi and xσk(i) and these products are accumulated in yσk(i) and yi, respectively. For the nonsymmetric matrix A defined by A = 10 −3 0 1 0 0 0 9 6 0 −2 0 3 0 8 7 0 0 0 6 0 7 5 4 0 0 0 0 9 13 0 0 0 0 5 −1 , (4.4) the 4 stripes of the matrix A stored in the rows of the val(:,:) array would be val(:,-1) 0 0 3 6 0 5 val(:, 0) 10 9 8 7 9 -1 val(:,+1) 0 -3 6 7 5 13 val(:,+2) 0 1 -2 0 4 0 . Jagged Diagonal Storage (JDS) The Jagged Diagonal Storage format can be useful for the implementation of iterative methods on parallel and vector processors (see Saad [184]). Like the Compressed Diagonal format, it gives a vector length essentially of the size of the matrix. It is more space-efficient than CDS at the cost of a gather/scatter operation. A simplified form of JDS, called ITPACK storage or Purdue storage, can be described as fol- lows. In the matrix from (4.4) all elements are shifted left: 10 −3 0 1 0 0 0 9 6 0 −2 0 3 0 8 7 0 0 0 6 0 7 5 4 0 0 0 0 9 13 0 0 0 0 5 −1 −→ 10 −3 1 9 6 −2 3 8 7 6 7 5 4 9 13 5 −1 after which the columns are stored consecutively. All rows are padded with zeros on the right to give them equal length. Corresponding to the array of matrix elements val(:,:), an array of column indices, col ind(:,:) is also stored: val(:, 1) 10 9 3 6 9 5 val(:, 2) −3 6 8 7 13 −1 val(:, 3) 1 −2 7 5 0 0 val(:, 4) 0 0 0 4 0 0 ,
- 70. 60 CHAPTER 4. RELATED ISSUES col ind(:, 1) 1 2 1 2 5 5 col ind(:, 2) 2 3 3 4 6 6 col ind(:, 3) 4 5 4 5 0 0 col ind(:, 4) 0 0 0 6 0 0 . It is clear that the padding zeros in this structure may be a disadvantage, especially if the bandwidth of the matrix varies strongly. Therefore, in the CRS format, we reorder the rows of the matrix decreasingly according to the number of nonzeros per row. The compressed and permuted diagonals are then stored in a linear array. The new data structure is called jagged diagonals. The number of jagged diagonals is equal to the number of nonzeros in the first row, i.e., the largest number of nonzeros in any row of A. The data structure to represent the n × n matrix A therefore consists of a permutation array (perm(1:n)) which reorders the rows, a floating-point array (jdiag(:)) containing the jagged diagonals in succession, an integer array (col ind(:)) containing the corresponding column indices indices, and finally a pointer array (jd ptr(:)) whose elements point to the beginning of each jagged diagonal. The advantages of JDS for matrix multiplications are discussed by Saad in [184]. The JDS format for the above matrix A in using the linear arrays {perm, jdiag, col ind, jd ptr} is given below (jagged diagonals are separated by semicolons) jdiag 1 3 7 8 10 2; 9 9 8 · · · -1; 9 6 7 5; 13 col ind 1 1 2 3 1 5; 4 2 3 · · · 6; 5 3 4 5; 6 perm 5 2 3 4 1 6 jd ptr 1 7 13 17 . Skyline Storage (SKS) The final storage scheme we consider is for skyline matrices, which are also called variable band or profile matrices (see Duff, Erisman and Reid [79]). It is mostly of importance in direct solution methods, but it can be used for handling the diagonal blocks in block matrix factorization methods. A major advantage of solving linear systems having skyline coefficient matrices is that when piv- oting is not necessary, the skyline structure is preserved during Gaussian elimination. If the matrix is symmetric, we only store its lower triangular part. A straightforward approach in storing the el- ements of a skyline matrix is to place all the rows (in order) into a floating-point array (val(:)), and then keep an integer array (row ptr(:)) whose elements point to the beginning of each row. The column indices of the nonzeros stored in val(:) are easily derived and are not stored. For a nonsymmetric skyline matrix such as the one illustrated in Figure 4.1, we store the lower triangular elements in SKS format, and store the upper triangular elements in a column-oriented SKS format (transpose stored in row-wise SKS format). These two separated substructures can be linked in a variety of ways. One approach, discussed by Saad in [185], is to store each row of the lower triangular part and each column of the upper triangular part contiguously into the floating- point array (val(:)). An additional pointer is then needed to determine where the diagonal elements, which separate the lower triangular elements from the upper triangular elements, are located. 4.3.2 Matrix vector products In many of the iterative methods discussed earlier, both the product of a matrix and that of its transpose times a vector are needed, that is, given an input vector x we want to compute products y = Ax and y = AT x. We will present these algorithms for two of the storage formats from §4.3: CRS and CDS.
- 71. 4.3. DATA STRUCTURES 61 + x x x + x x x x + x x x x + x x x x x + x x x x x x + x x x x + x x x x x x + x x x x x x x x + x x x x + x x x x x x x + x x x x + x x x x x + x x x x x + x x + x x x x x + x + Figure 4.1: Profile of a nonsymmetric skyline or variable-band matrix. CRS Matrix-Vector Product The matrix vector product y = Ax using CRS format can be expressed in the usual way: yi = j ai,jxj, since this traverses the rows of the matrix A. For an n×n matrix A, the matrix-vector multiplication is given by for i = 1, n y(i) = 0 for j = row_ptr(i), row_ptr(i+1) - 1 y(i) = y(i) + val(j) * x(col_ind(j)) end; end; Since this method only multiplies nonzero matrix entries, the operation count is 2 times the number of nonzero elements in A, which is a significant savings over the dense operation requirement of 2n2 . For the transpose product y = AT x we cannot use the equation yi = j (AT )i,jxj = j aj,ixj, since this implies traversing columns of the matrix, an extremely inefficient operation for matrices stored in CRS format. Hence, we switch indices to for all j, do for all i: yi ← yi + aj,ixj. The matrix-vector multiplication involving AT is then given by for i = 1, n y(i) = 0 end; for j = 1, n
- 72. 62 CHAPTER 4. RELATED ISSUES for i = row_ptr(j), row_ptr(j+1)-1 y(col_ind(i)) = y(col_ind(i)) + val(i) * x(j) end; end; Both matrix-vector products above have largely the same structure, and both use indirect ad- dressing. Hence, their vectorizability properties are the same on any given computer. However, the first product (y = Ax) has a more favorable memory access pattern in that (per iteration of the outer loop) it reads two vectors of data (a row of matrix A and the input vector x) and writes one scalar. The transpose product (y = AT x) on the other hand reads one element of the input vector, one row of matrix A, and both reads and writes the result vector y. Unless the machine on which these methods are implemented has three separate memory paths (e.g., Cray Y-MP), the memory traffic will then limit the performance. This is an important consideration for RISC-based architectures. CDS Matrix-Vector Product If the n×n matrix A is stored in CDS format, it is still possible to perform a matrix-vector product y = Ax by either rows or columns, but this does not take advantage of the CDS format. The idea is to make a change in coordinates in the doubly-nested loop. Replacing j → i + j we get yi ← yi + ai,jxj ⇒ yi ← yi + ai,i+jxi+j . With the index i in the inner loop we see that the expression ai,i+j accesses the jth diagonal of the matrix (where the main diagonal has number 0). The algorithm will now have a doubly-nested loop with the outer loop enumerating the diago- nals diag=-p,q with p and q the (nonnegative) numbers of diagonals to the left and right of the main diagonal. The bounds for the inner loop follow from the requirement that 1 ≤ i, i + j ≤ n. The algorithm becomes for i = 1, n y(i) = 0 end; for diag = -diag_left, diag_right for loc = max(1,1-diag), min(n,n-diag) y(loc) = y(loc) + val(loc,diag) * x(loc+diag) end; end; The transpose matrix-vector product y = AT x is a minor variation of the algorithm above. Using the update formula yi ← yi + ai+j,ixj = yi + ai+j,i+j−jxi+j we obtain for i = 1, n y(i) = 0 end; for diag = -diag_right, diag_left for loc = max(1,1-diag), min(n,n-diag) y(loc) = y(loc) + val(loc+diag, -diag) * x(loc+diag) end; end;
- 73. 4.3. DATA STRUCTURES 63 The memory access for the CDS-based matrix-vector product y = Ax (or y = AT x) is three vectors per inner iteration. On the other hand, there is no indirect addressing, and the algorithm is vectorizable with vector lengths of essentially the matrix order n. Because of the regular data access, most machines can perform this algorithm efficiently by keeping three base registers and using simple offset addressing. 4.3.3 Sparse Incomplete Factorizations Efficient preconditioners for iterative methods can be found by performing an incomplete factor- ization of the coefficient matrix. In this section, we discuss the incomplete factorization of an n×n matrix A stored in the CRS format, and routines to solve a system with such a factorization. At first we only consider a factorization of the D-ILU type, that is, the simplest type of factorization in which no “fill” is allowed, even if the matrix has a nonzero in the fill position (see section 3.4.2). Later we will consider factorizations that allow higher levels of fill. Such factorizations consider- ably more complicated to code, but they are essential for complicated differential equations. The solution routines are applicable in both cases. For iterative methods, such as QMR, that involve a transpose matrix vector product we need to consider solving a system with the transpose of as factorization as well. Generating a CRS-based D-ILU Incomplete Factorization In this subsection we will consider a matrix split as A = DA + LA + UA in diagonal, lower and upper triangular part, and an incomplete factorization preconditioner of the form (DA + LA)D−1 A (DA + UA). In this way, we only need to store a diagonal matrix D containing the pivots of the factorization. Hence,it suffices to allocate for the preconditioner only a pivot array of length n (pivots(1:n)). In fact, we will store the inverses of the pivots rather than the pivots themselves. This implies that during the system solution no divisions have to be performed. Additionally, we assume that an extra integer array diag ptr(1:n) has been allo- cated that contains the column (or row) indices of the diagonal elements in each row, that is, val(diag ptr(i)) = ai,i. The factorization begins by copying the matrix diagonal for i = 1, n pivots(i) = val(diag_ptr(i)) end; Each elimination step starts by inverting the pivot for i = 1, n pivots(i) = 1 / pivots(i) For all nonzero elements ai,j with j > i, we next check whether aj,i is a nonzero matrix element, since this is the only element that can cause fill with ai,j. for j = diag_ptr(i)+1, row_ptr(i+1)-1 found = FALSE for k = row_ptr(col_ind(j)), diag_ptr(col_ind(j))-1 if(col_ind(k) = i) then found = TRUE element = val(k) endif end;
- 74. 64 CHAPTER 4. RELATED ISSUES If so, we update aj,j. if (found = TRUE) val(diag_ptr(col_ind(j))) = val(diag_ptr(col_ind(j))) - element * pivots(i) * val(j) end; end; CRS-based Factorization Solve The system LUy = x can be solved in the usual manner by introducing a temporary vector z: Lz = x, Uy = z. We have a choice between several equivalent ways of solving the system: LU = (D + LA)D−1 (D + UA) = (I + LAD−1 )(D + UA) = (D + LA)(I + D−1 UA) = (I + LAD−1 )D(I + D−1 UA) The first and fourth formulae are not suitable since they require both multiplication and division with D; the difference between the second and third is only one of ease of coding. In this section we use the third formula; in the next section we will use the second for the transpose system solution. Both halves of the solution have largely the same structure as the matrix vector multiplication. for i = 1, n sum = 0 for j = row_ptr(i), diag_ptr(i)-1 sum = sum + val(j) * z(col_ind(j)) end; z(i) = pivots(i) * (x(i)-sum) end; for i = n, 1, (step -1) sum = 0 for j = diag(i)+1, row_ptr(i+1)-1 sum = sum + val(j) * y(col_ind(j)) y(i) = z(i) - pivots(i) * sum end; end; The temporary vector z can be eliminated by reusing the space for y; algorithmically, z can even overwrite x, but overwriting input data is in general not recommended. CRS-based Factorization Transpose Solve Solving the transpose system (LU)T y = x is slightly more involved. In the usual formulation we traverse rows when solving a factored system, but here we can only access columns of the matrices LT and UT (at less than prohibitive cost). The key idea is to distribute each newly computed component of a triangular solve immediately over the remaining right-hand-side. For instance, if we write a lower triangular matrix as L = (l∗1, l∗2, . . . ), then the system Ly = x can be written as x = l∗1y1 + l∗2y2 + · · ·. Hence, after computing y1 we modify x ← x − l∗1y1, and so on. Upper triangular systems are treated in a similar manner. With this algorithm we only access columns of the triangular systems. Solving a transpose system with a matrix stored in CRS format essentially means that we access rows of L and U. The algorithm now becomes
- 75. 4.3. DATA STRUCTURES 65 for i = 1, n x_tmp(i) = x(i) end; for i = 1, n z(i) = x_tmp(i) tmp = pivots(i) * z(i) for j = diag_ptr(i)+1, row_ptr(i+1)-1 x_tmp(col_ind(j)) = x_tmp(col_ind(j)) - tmp * val(j) end; end; for i = n, 1 (step -1) y(i) = pivots(i) * z(i) for j = row_ptr(i), diag_ptr(i)-1 z(col_ind(j)) = z(col_ind(j)) - val(j) * y(i) end; end; The extra temporary x tmp is used only for clarity, and can be overlapped with z. Both x tmp and z can be considered to be equivalent to y. Overall, a CRS-based preconditioner solve uses short vector lengths, indirect addressing, and has essentially the same memory traffic patterns as that of the matrix-vector product. Generating a CRS-based ILU(k) Incomplete Factorization Incomplete factorizations with several levels of fill allowed are more accurate than the D-ILU factorization described above. On the other hand, they require more storage, and are considerably harder to implement (much of this section is based on algorithms for a full factorization of a sparse matrix as found in Duff, Erisman and Reid [79]). As a preliminary, we need an algorithm for adding two vectors x and y, both stored in sparse storage. Let lx be the number of nonzero components in x, let x be stored in x, and let xind be an integer array such that if xind(j)=i then x(j) = xi. Similarly, y is stored as ly, y, yind. We now add x ← x + y by first copying y into a full vector w then adding w to x. The total number of operations will be O(lx + ly)3 : % copy y into w for i=1,ly w( yind(i) ) = y(i) % add w to x wherever x is already nonzero for i=1,lx if w( xind(i) ) <> 0 x(i) = x(i) + w( xind(i) ) w( xind(i) ) = 0 % add w to x by creating new components % wherever x is still zero for i=1,ly if w( yind(i) ) <> 0 then lx = lx+1 3This is not counting the initial zeroing of the w array.
- 76. 66 CHAPTER 4. RELATED ISSUES xind(lx) = yind(i) x(lx) = w( yind(i) ) endif In order to add a sequence of vectors x ← x + k y(k) , we add the y(i) vectors into w before executing the writes into x. A different implementation would be possible, where w is allocated as a sparse vector and its sparsity pattern is constructed during the additions. We will not discuss this possibility any further. For a slight refinement of the above algorithm, we will add levels to the nonzero components: we assume integer vectors xlev and ylev of length lx and ly respectively, and a full length level vector wlev corresponding to w. The addition algorithm then becomes: % copy y into w for i=1,ly w( yind(i) ) = y(i) wlev( yind(i) ) = ylev(i) % add w to x wherever x is already nonzero; % don’t change the levels for i=1,lx if w( xind(i) ) <> 0 x(i) = x(i) + w( xind(i) ) w( xind(i) ) = 0 % add w to x by creating new components % wherever x is still zero; % carry over levels for i=1,ly if w( yind(i) ) <> 0 then lx = lx+1 x(lx) = w( yind(i) ) xind(lx) = yind(i) xlev(lx) = wlev( yind(i) ) endif We can now describe the ILU(k) factorization. The algorithm starts out with the matrix A, and gradually builds up a factorization M of the form M = (D + L)(I + D−1 U), where L, D−1 , and D−1 U are stored in the lower triangle, diagonal and upper triangle of the array M respectively. The particular form of the factorization is chosen to minimize the number of times that the full vector w is copied back to sparse form. Specifically, we use a sparse form of the following factorization scheme: for k=1,n for j=1,k-1 for i=j+1,n a(k,i) = a(k,i) - a(k,j)*a(j,i) for j=k+1,n a(k,j) = a(k,j)/a(k,k) This is a row-oriented version of the traditional ‘left-looking’ factorization algorithm. We will describe an incomplete factorization that controls fill-in through levels (see equa- tion (3.1)). Alternatively we could use a drop tolerance (section 3.4.2), but this is less attractive from a point of implementation. With fill levels we can perform the factorization symbolically at first, determining storage demands and reusing this information through a number of linear systems of the same sparsity structure. Such preprocessing and reuse of information is not possible with fill controlled by a drop tolerance criterium.
- 77. 4.4. PARALLELISM 67 The matrix arrays A and M are assumed to be in compressed row storage, with no particular ordering of the elements inside each row, but arrays adiag and mdiag point to the locations of the diagonal elements. for row=1,n % go through elements A(row,col) with col<row COPY ROW row OF A() INTO DENSE VECTOR w for col=aptr(row),aptr(row+1)-1 if aind(col) < row then acol = aind(col) MULTIPLY ROW acol OF M() BY A(col) SUBTRACT THE RESULT FROM w ALLOWING FILL-IN UP TO LEVEL k endif INSERT w IN ROW row OF M() % invert the pivot M(mdiag(row)) = 1/M(mdiag(row)) % normalize the row of U for col=mptr(row),mptr(row+1)-1 if mind(col) > row M(col) = M(col) * M(mdiag(row)) The structure of a particular sparse matrix is likely to apply to a sequence of problems, for instance on different time-steps, or during a Newton iteration. Thus it may pay off to perform the above incomplete factorization first symbolically to determine the amount and location of fill-in and use this structure for the numerically different but structurally identical matrices. In this case, the array for the numerical values can be used to store the levels during the symbolic factorization phase. 4.4 Parallelism In this section we discuss aspects of parallelism in the iterative methods discussed in this book. Since the iterative methods share most of their computational kernels we will discuss these independent of the method. The basic time-consuming kernels of iterative schemes are: • inner products, • vector updates, • matrix–vector products, e.g., Ap(i) (for some methods also AT p(i) ), • preconditioner solves. We will examine each of these in turn. We will conclude this section by discussing two par- ticular issues, namely computational wavefronts in the SOR method, and block operations in the GMRES method. 4.4.1 Inner products The computation of an inner product of two vectors can be easily parallelized; each processor computes the inner product of corresponding segments of each vector (local inner products or LIPs). On distributed-memory machines the LIPs then have to be sent to other processors to be combined for the global inner product. This can be done either with an all-to-all send where every
- 78. 68 CHAPTER 4. RELATED ISSUES processor performs the summation of the LIPs, or by a global accumulation in one processor, followed by a broadcast of the final result. Clearly, this step requires communication. For shared-memory machines, the accumulation of LIPs can be implemented as a critical sec- tion where all processors add their local result in turn to the global result, or as a piece of serial code, where one processor performs the summations. Overlapping communication and computation Clearly, in the usual formulation of conjugate gradient-type methods the inner products induce a synchronization of the processors, since they cannot progress until the final result has been com- puted: updating x(i+1) and r(i+1) can only begin after completing the inner product for αi. Since on a distributed-memory machine communication is needed for the inner product, we cannot over- lap this communication with useful computation. The same observation applies to updating p(i) , which can only begin after completing the inner product for βi−1. Figure 4.2 shows a variant of CG, in which all communication time may be overlapped with useful computations. This is just a reorganized version of the original CG scheme, and is therefore precisely as stable. Another advantage over other approaches (see below) is that no additional operations are required. This rearrangement is based on two tricks. The first is that updating the iterate is delayed to mask the communication stage of the p(i)T Ap(i) inner product. The second trick relies on splitting the (symmetric) preconditioner as M = LLT , so one first computes L−1 r(i) , after which the inner product r(i)T M−1 r(i) can be computed as sT s where s = L−1 r(i) . The computation of L−T s will then mask the communication stage of the inner product. x(−1) = x(0) = initial guess; r(0) = b − Ax(0) ; p(−1) = 0; β−1 = 0; α−1 = 0; s = L−1 r(0) ; ρ0 = (s, s) for i = 0, 1, 2, .... w(i) = L−T s; p(i) = w(i) + βi−1p(i−1) ; q(i) = Ap(i) ; γ = (p(i) , q(i) ); x(i) = x(i−1) + αi−1p(i−1) ; αi = ρi/γ; r(i+1) = r(i) − αiq(i) ; s = L−1 r(i+1) ; ρi+1 = (s, s); if r(i+1) small enough then x(i+1) = x(i) + αip(i) quit; endif βi = ρi+1/ρi; end; Figure 4.2: A rearrangement of Conjugate Gradient for parallelism Under the assumptions that we have made, CG can be efficiently parallelized as follows:
- 79. 4.4. PARALLELISM 69 1. The communication required for the reduction of the inner product for γ can be overlapped with the update for x(i) , (which could in fact have been done in the previous iteration step). 2. The reduction of the inner product for ρi+1 can be overlapped with the remaining part of the preconditioning operation at the beginning of the next iteration. 3. The computation of a segment of p(i) can be followed immediately by the computation of a segment of q(i) , and this can be followed by the computation of a part of the inner product. This saves on load operations for segments of p(i) and q(i) . For a more detailed discussion see Demmel, Heath and Van der Vorst [66]. This algorithm can be extended trivially to preconditioners of LDLT form, and nonsymmetric preconditioners in the Biconjugate Gradient Method. Fewer synchronization points Several authors have found ways to eliminate some of the synchronization points induced by the inner products in methods such as CG. One strategy has been to replace one of the two inner prod- ucts typically present in conjugate gradient-like methods by one or two others in such a way that all inner products can be performed simultaneously. The global communication can then be pack- aged. A first such method was proposed by Saad [181] with a modification to improve its stability suggested by Meurant [155]. Recently, related methods have been proposed by Chronopoulos and Gear [54], D’Azevedo and Romine [61], and Eijkhout [87]. These schemes can also be applied to nonsymmetric methods such as BiCG. The stability of such methods is discussed by D’Azevedo, Eijkhout and Romine [60]. Another approach is to generate a number of successive Krylov vectors (see §2.3.4) and orthog- onalize these as a block (see Van Rosendale [208], and Chronopoulos and Gear [54]). 4.4.2 Vector updates Vector updates are trivially parallelizable: each processor updates its own segment. 4.4.3 Matrix-vector products The matrix–vector products are often easily parallelized on shared-memory machines by split- ting the matrix in strips corresponding to the vector segments. Each processor then computes the matrix–vector product of one strip. For distributed-memory machines, there may be a problem if each processor has only a segment of the vector in its memory. Depending on the bandwidth of the matrix, we may need communication for other elements of the vector, which may lead to com- munication bottlenecks. However, many sparse matrix problems arise from a network in which only nearby nodes are connected. For example, matrices stemming from finite difference or finite element problems typically involve only local connections: matrix element ai,j is nonzero only if variables i and j are physically close. In such a case, it seems natural to subdivide the network, or grid, into suitable blocks and to distribute them over the processors. When computing Api, each processor requires the values of pi at some nodes in neighboring blocks. If the number of con- nections to these neighboring blocks is small compared to the number of internal nodes, then the communication time can be overlapped with computational work. For more detailed discussions on implementation aspects for distributed memory systems, see De Sturler [62] and Pommerell [174]. 4.4.4 Preconditioning Preconditioning is often the most problematic part of parallelizing an iterative method. We will mention a number of approaches to obtaining parallelism in preconditioning.
- 80. 70 CHAPTER 4. RELATED ISSUES Discovering parallelism in sequential preconditioners. Certain preconditioners were not de- veloped with parallelism in mind, but they can be executed in parallel. Some examples are domain decomposition methods (see §5.4), which provide a high degree of coarse grained parallelism, and polynomial preconditioners (see §3.5), which have the same parallelism as the matrix-vector prod- uct. Incomplete factorization preconditioners are usually much harder to parallelize: using wave- fronts of independent computations (see for instance Paolini and Radicati di Brozolo [169]) a modest amount of parallelism can be attained, but the implementation is complicated. For in- stance, a central difference discretization on regular grids gives wavefronts that are hyperplanes (see Van der Vorst [201, 203]). More parallel variants of sequential preconditioners. Variants of existing sequential incom- plete factorization preconditioners with a higher degree of parallelism have been devised, though they are perhaps less efficient in purely scalar terms than their ancestors. Some examples are: re- orderings of the variables (see Duff and Meurant [78] and Eijkhout [84]), expansion of the factors in a truncated Neumann series (see Van der Vorst [199]), various block factorization methods (see Axelsson and Eijkhout [15] and Axelsson and Polman [21]), and multicolor preconditioners. Multicolor preconditioners have optimal parallelism among incomplete factorization methods, since the minimal number of sequential steps equals the color number of the matrix graphs. For theory and appplications to parallelism see Jones and Plassman [126, 127]. Fully decoupled preconditioners. If all processors execute their part of the preconditioner solve without further communication, the overall method is technically a block Jacobi preconditioner (see §3.2.1). While their parallel execution is very efficient, they may not be as effective as more complicated, less parallel preconditioners, since improvement in the number of iterations may be only modest. To get a bigger improvement while retaining the efficient parallel execution, Radi- cati di Brozolo and Robert [177] suggest that one construct incomplete decompositions on slightly overlapping domains. This requires communication similar to that for matrix–vector products. 4.4.5 Wavefronts in the Gauss-Seidel and Conjugate Gradient methods At first sight, the Gauss-Seidel method (and the SOR method which has the same basic structure) seems to be a fully sequential method. A more careful analysis, however, reveals a high degree of parallelism if the method is applied to sparse matrices such as those arising from discretized partial differential equations. We start by partitioning the unknowns in wavefronts. The first wavefront contains those un- knowns that (in the directed graph of D − L) have no predecessor; subsequent wavefronts are then sets (this definition is not necessarily unique) of successors of elements of the previous wave- front(s), such that no successor/predecessor relations hold among the elements of this set. It is clear that all elements of a wavefront can be processed simultaneously, so the sequential time of solving a system with D − L can be reduced to the number of wavefronts. Next, we observe that the unknowns in a wavefront can be computed as soon as all wavefronts containing its predecessors have been computed. Thus we can, in the absence of tests for con- vergence, have components from several iterations being computed simultaneously. Adams and Jordan [2] observe that in this way the natural ordering of unknowns gives an iterative method that is mathematically equivalent to a multi-color ordering. In the multi-color ordering, all wavefronts of the same color are processed simultaneously. This reduces the number of sequential steps for solving the Gauss-Seidel matrix to the number of colors, which is the smallest number d such that wavefront i contains no elements that are a predecessor of an element in wavefront i + d.
- 81. 4.4. PARALLELISM 71 As demonstrated by O’Leary [163], SOR theory still holds in an approximate sense for multi- colored matrices. The above observation that the Gauss-Seidel method with the natural ordering is equivalent to a multicoloring cannot be extended to the SSOR method or wavefront-based in- complete factorization preconditioners for the Conjugate Gradient method. In fact, tests by Duff and Meurant [78] and an analysis by Eijkhout [84] show that multicolor incomplete factorization preconditioners in general may take a considerably larger number of iterations to converge than preconditioners based on the natural ordering. Whether this is offset by the increased parallelism depends on the application and the computer architecture. 4.4.6 Blocked operations in the GMRES method In addition to the usual matrix-vector product, inner products and vector updates, the precondi- tioned GMRES method (see §2.3.4) has a kernel where one new vector, M−1 Av(j) , is orthogonal- ized against the previously built orthogonal set {v(1) , v(2) ,..., v(j) }. In our version, this is done using Level 1 BLAS, which may be quite inefficient. To incorporate Level 2 BLAS we can apply either Householder orthogonalization or classical Gram-Schmidt twice (which mitigates classical Gram-Schmidt’s potential instability; see Saad [184]). Both approaches significantly increase the computational work, but using classical Gram-Schmidt has the advantage that all inner products can be performed simultaneously; that is, their communication can be packaged. This may increase the efficiency of the computation significantly. Another way to obtain more parallelism and data locality is to generate a basis {v(1) , Av(1) , ..., Am v(1) } for the Krylov subspace first, and to orthogonalize this set afterwards; this is called m-step GMRES(m) (see Kim and Chronopoulos [138]). (Compare this to the GMRES method in §2.3.4, where each new vector is immediately orthogonalized to all previous vectors.) This approach does not increase the computational work and, in contrast to CG, the numerical instability due to generating a possibly near-dependent set is not necessarily a drawback.
- 82. 72 CHAPTER 4. RELATED ISSUES
- 83. Chapter 5 Remaining topics 5.1 The Lanczos Connection As discussed by Paige and Saunders in [167] and by Golub and Van Loan in [108], it is straightfor- ward to derive the conjugate gradient method for solving symmetric positive definite linear systems from the Lanczos algorithm for solving symmetric eigensystems and vice versa. As an example, let us consider how one can derive the Lanczos process for symmetric eigensystems from the (un- preconditioned) conjugate gradient method. Suppose we define the n × k matrix R(k) by Rk = [r(0) , r(1) , . . . , r(k−1) ], and the k × k upper bidiagonal matrix Bk by Bk = 1 −β1 · · · 0 1 −β2 ... ... ... ... ... ... ... −βk−1 0 · · · 1 , where the sequences {r(k) } and {βk} are defined by the standard conjugate gradient algorithm discussed in §2.3.1. From the equations p(j) = r(j−1) + βj−1p(j−1) , j = 2, 3, . . . , k , and p(1) = r(0) , we have Rk = PkBk, where Pk = [p(1) , p(2) , . . . , p(k) ]. Assuming the elements of the sequence {p(j) } are A-conjugate, it follows that ˆTk = RT k ARk = BT k ˆΛkBk is a tridiagonal matrix since ˆΛk = p(1)T Ap(1) 0 · · · 0 0 p(2)T Ap(2) ... ... ... ... ... ... ... 0 0 · · · 0 p(k)T Ap(k) . 73
- 84. 74 CHAPTER 5. REMAINING TOPICS Since span{p(1) , p(2) , . . . , p(j) } = span{r(0) , r(1) , . . . , r(j−1) } and since the elements of {r(j) } are mutually orthogonal, it can be shown that the columns of n × k matrix Qk = Rk∆−1 form an orthonormal basis for the subspace span{b, Ab, . . . , Ak−1 b}, where ∆ is a diagonal matrix whose ith diagonal element is r(i) 2. The columns of the matrix Qk are the Lanczos vectors (see Parlett [170]) whose associated projection of A is the tridiagonal matrix Tk = ∆−1 BT k ˆΛkBk∆−1 . (5.1) The extremal eigenvalues of Tk approximate those of the matrix A. Hence, the diagonal and sub- diagonal elements of Tk in (5.1), which are readily available during iterations of the conjugate gradient algorithm (§2.3.1), can be used to construct Tk after k CG iterations. This allows us to obtain good approximations to the extremal eigenvalues (and hence the condition number) of the matrix A while we are generating approximations, x(k) , to the solution of the linear system Ax = b. For a nonsymmetric matrix A, an equivalent nonsymmetric Lanczos algorithm (see Lanc- zos [141]) would produce a nonsymmetric matrix Tk in (5.1) whose extremal eigenvalues (which may include complex-conjugate pairs) approximate those of A. The nonsymmetric Lanczos method is equivalent to the BiCG method discussed in §2.3.5. 5.2 Block and s-step Iterative Methods The methods discussed so far are all subspace methods, that is, in every iteration they extend the dimension of the subspace generated. In fact, they generate an orthogonal basis for this subspace, by orthogonalizing the newly generated vector with respect to the previous basis vectors. However, in the case of nonsymmetric coefficient matrices the newly generated vector may be almost linearly dependent on the existing basis. To prevent break-down or severe numerical error in such instances, methods have been proposed that perform a look-ahead step (see Freund, Gutknecht and Nachtigal [100], Parlett, Taylor and Liu [171], and Freund and Nachtigal [101]). Several new, unorthogonalized, basis vectors are generated and are then orthogonalized with respect to the subspace already generated. Instead of generating a basis, such a method generates a series of low-dimensional orthogonal subspaces. The s-step iterative methods of Chronopoulos and Gear [54] use this strategy of generating unorthogonalized vectors and processing them as a block to reduce computational overhead and improve processor cache behaviour. If conjugate gradient methods are considered to generate a factorization of a tridiagonal re- duction of the original matrix, then look-ahead methods generate a block factorization of a block tridiagonal reduction of the matrix. A block tridiagonal reduction is also effected by the Block Lanczos algorithm and the Block Conjugate Gradient method (see O’Leary [162]). Such methods operate on multiple linear systems with the same coefficient matrix simultaneously, for instance with multiple right hand sides, or the same right hand side but with different initial guesses. Since these block methods use multiple search directions in each step, their convergence behavior is better than for ordinary methods. In fact, one can show that the spectrum of the matrix is effectively reduced by the nb − 1 smallest eigenvalues, where nb is the block size. 5.3 Reduced System Preconditioning As we have seen earlier, a suitable preconditioner for CG is a matrix M such that the system M−1 Ax = M−1 f requires fewer iterations to solve than Ax = f does, and for which systems Mz = r can be solved efficiently. The first property is independent of the machine used, while the second is highly ma- chine dependent. Choosing the best preconditioner involves balancing those two criteria in a way
- 85. 5.4. DOMAIN DECOMPOSITION METHODS 75 that minimizes the overall computation time. One balancing approach used for matrices A arising from 5-point finite difference discretization of second order elliptic partial differential equations (PDEs) with Dirichlet boundary conditions involves solving a reduced system. Specifically, for an n × n grid, we can use a point red-black ordering of the nodes to get Ax = DR C CT DB xR xB = fR fB , (5.2) where DR and DB are diagonal, and C is a well-structured sparse matrix with 5 nonzero diagonals if n is even and 4 nonzero diagonals if n is odd. Applying one step of block Gaussian elimination (or computing the Schur complement; see Golub and Van Loan [108]) we have I O −CT D−1 R I DR C CT DB xR xB = I O −CT D−1 R I fR fB which reduces to DR C O DB − CT DR −1 C xR xB = fR fB − CT DR −1 fR . With proper scaling (left and right multiplication by DB −1/2 ), we obtain from the second block equation the reduced system (I − HT H)y = g , (5.3) where H = D −1/2 R CD −1/2 B , y = D 1/2 B xB, and g = D −1/2 B (fB − CT D−1 R fR). The linear system (5.3) is of order n2 /2 for even n and of order (n2 − 1)/2 for odd n. Once y is determined, the solution x is easily retrieved from y. The values on the black points are those that would be obtained from a red/black ordered SSOR preconditioner on the full system, so we expect faster convergence. The number of nonzero coefficients is reduced, although the coefficient matrix in (5.3) has nine nonzero diagonals. Therefore it has higher density and offers more data locality. Meier and Sameh [149] demonstrate that the reduced system approach on hierarchical memory machines such as the Alliant FX/8 is over 3 times faster than unpreconditioned CG for Poisson’s equation on n×n grids with n ≥ 250. For 3-dimensional elliptic PDEs, the reduced system approach yields a block tridiagonal ma- trix C in (5.2) having diagonal blocks of the structure of C from the 2-dimensional case and off- diagonal blocks that are diagonal matrices. Computing the reduced system explicitly leads to an unreasonable increase in the computational complexity of solving Ax = f. The matrix products required to solve (5.3) would therefore be performed implicitly which could significantly decrease performance. However, as Meier and Sameh show [149], the reduced system approach can still be about 2-3 times as fast as the conjugate gradient method with Jacobi preconditioning for 3- dimensional problems. 5.4 Domain Decomposition Methods In recent years, much attention has been given to domain decomposition methods for linear el- liptic problems that are based on a partitioning of the domain of the physical problem. Since the subdomains can be handled independently, such methods are very attractive for coarse-grain par- allel computers. On the other hand, it should be stressed that they can be very effective even on sequential computers. In this brief survey, we shall restrict ourselves to the standard second order self-adjoint scalar elliptic problems in two dimensions of the form: − · (a(x, y) u) = f(x, y) (5.4)
- 86. 76 CHAPTER 5. REMAINING TOPICS where a(x, y) is a positive function on the domain Ω, on whose boundary the value of u is pre- scribed (the Dirichlet problem). For more general problems, and a good set of references, the reader is referred to the series of proceedings [47, 48, 49, 106, 134, 176]. For the discretization of (5.4), we shall assume for simplicity that Ω is triangulated by a set TH of nonoverlapping coarse triangles (subdomains) Ωi, i = 1, ..., p with nH internal vertices. The Ωi’s are in turn further refined into a set of smaller triangles Th with n internal vertices in total. Here H, h denote the coarse and fine mesh size respectively. By a standard Ritz-Galerkin method using piecewise linear triangular basis elements on (5.4), we obtain an n × n symmetric positive definite linear system Au = f. Generally, there are two kinds of approaches depending on whether the subdomains overlap with one another (Schwarz methods) or are separated from one another by interfaces (Schur Com- plement methods, iterative substructuring). We shall present domain decomposition methods as preconditioners for the linear system Au = f or to a reduced (Schur Complement) system SBuB = gB defined on the interfaces in the non- overlapping formulation. When used with the standard Krylov subspace methods discussed else- where in this book, the user has to supply a procedure for computing Av or Sw given v or w and the algorithms to be described herein computes K−1 v. The computation of Av is a simple sparse matrix-vector multiply, but Sw may require subdomain solves, as will be described later. 5.4.1 Overlapping Subdomain Methods In this approach, each substructure Ωi is extended to a larger substructure Ωi containing ni internal vertices and all the triangles Ti ⊂ Th, within a distance δ from Ωi, where δ refers to the amount of overlap. Let Ai, AH denote the the discretizations of (5.4) on the subdomain triangulation Ti and the coarse triangulation TH respectively. Let RT i denote the extension operator which extends (by zero) a function on Ti to Th and Ri the corresponding pointwise restriction operator. Similarly, let RT H denote the interpolation operator which maps a function on the coarse grid TH onto the fine grid Th by piecewise linear interpolation and RH the corresponding weighted restriction operator. With these notations, the Additive Schwarz Preconditioner Kas for the system Au = f can be compactly described as: K−1 as v = RT HA−1 H RHv + p i=1 RT i A −1 i Riv. Note that the right hand side can be computed using p subdomain solves using the Ai’s, plus a coarse grid solve using AH, all of which can be computed in parallel. Each term RT i A −1 i Riv should be evaluated in three steps: (1) Restriction: vi = Riv, (2) Subdomain solves for wi: Aiwi = vi, (3) Interpolation: yi = RT i wi. The coarse grid solve is handled in the same manner. The theory of Dryja and Widlund [75] shows that the condition number of K−1 as A is bounded independently of both the coarse grid size H and the fine grid size h, provided there is “sufficient” overlap between Ωi and Ωi (essentially it means that the ratio δ/H of the distance δ of the boundary ∂Ωi to ∂Ωi should be uniformly bounded from below as h → 0.) If the coarse grid solve term is left out, then the condition number grows as O(1/H2 ), reflecting the lack of global coupling provided by the coarse grid. For the purpose of implementations, it is often useful to interpret the definition of Kas in matrix notation. Thus the decomposition of Ω into Ωi’s corresponds to partitioning of the components of the vector u into p overlapping groups of index sets Ii’s, each with ni components. The ni × ni matrix Ai is simply a principal submatrix of A corresponding to the index set Ii. RT i is a n × ni matrix defined by its action on a vector ui defined on Ti as: (RT i ui)j = (ui)j if j ∈ Ii but is zero otherwise. Similarly, the action of its transpose Riu forms an ni-vector by picking off
- 87. 5.4. DOMAIN DECOMPOSITION METHODS 77 the components of u corresponding to Ii. Analogously, RT H is an n × nH matrix with entries corresponding to piecewise linear interpolation and its transpose can be interpreted as a weighted restriction matrix. If Th is obtained from TH by nested refinement, the action of RH, RT H can be efficiently computed as in a standard multigrid algorithm. Note that the matrices RT i , Ri, RT H, RH are defined by their actions and need not be stored explicitly. We also note that in this algebraic formulation, the preconditioner Kas can be extended to any matrix A, not necessarily one arising from a discretization of an elliptic problem. Once we have the partitioning index sets Ii’s, the matrices Ri, Ai are defined. Furthermore, if A is positive definite, then Ai is guaranteed to be nonsingular. The difficulty is in defining the “coarse grid” matrices AH, RH, which inherently depends on knowledge of the grid structure. This part of the preconditioner can be left out, at the expense of a deteriorating convergence rate as p increases. Radicati and Robert [177] have experimented with such an algebraic overlapping block Jacobi preconditioner. 5.4.2 Non-overlapping Subdomain Methods The easiest way to describe this approach is through matrix notation. The set of vertices of Th can be divided into two groups. The set of interior vertices I of all Ωi and the set of vertices B which lies on the boundaries i ∂Ωi of the coarse triangles Ωi in TH. We shall re-order u and f as u ≡ (uI, uB)T and f ≡ (fI, fB)T corresponding to this partition. In this ordering, equation (5.4) can be written as follows: AI AIB AT IB AB uI uB = fI fB . (5.5) We note that since the subdomains are uncoupled by the boundary vertices, AI = blockdiagonal(Ai) is block-diagonal with each block Ai being the stiffness matrix corresponding to the unknowns belonging to the interior vertices of subdomain Ωi. By a block LU-factorization of A, the system in (5.5) can be written as: I 0 AT IBA−1 I I AI AIB 0 SB uI uB = fI fB , (5.6) where SB ≡ AB − AT IBA−1 I AIB is the Schur complement of AB in A. By eliminating uI in (5.6), we arrive at the following equation for uB: SBuB = gB ≡ fB − AIBA−1 I fI. (5.7) We note the following properties of this Schur Complement system: 1. SB inherits the symmetric positive definiteness of A. 2. SB is dense in general and computing it explicitly requires as many solves on each subdo- main as there are points on each of its edges. 3. The condition number of SB is O(h−1 ), an improvement over the O(h−2 ) growth for A. 4. Given a vector vB defined on the boundary vertices B of TH, the matrix-vector product SBvB can be computed according to ABvB − AT IB(A−1 I (AIBvB)) where A−1 I involves p independent subdomain solves using A−1 i . 5. The right hand side gB can also be computed using p independent subdomain solves.
- 88. 78 CHAPTER 5. REMAINING TOPICS These properties make it possible to apply a preconditioned iterative method to (5.7), which is the basic method in the nonoverlapping substructuring approach. We will also need some good preconditioners to further improve the convergence of the Schur system. We shall first describe the Bramble-Pasciak-Schatz preconditioner [36]. For this, we need to further decompose B into two non-overlapping index sets: B = E ∪ VH (5.8) where VH = k Vk denote the set of nodes corresponding to the vertices Vk’s of TH, and E = i Ei denote the set of nodes on the edges Ei’s of the coarse triangles in TH, excluding the vertices belonging to VH. In addition to the coarse grid interpolation and restriction operators RH, RT H defined before, we shall also need a new set of interpolation and restriction operators for the edges Ei’s. Let REi be the pointwise restriction operator (an nEi × n matrix, where nEi is the number of vertices on the edge Ei) onto the edge Ei defined by its action (REi uB)j = (uB)j if j ∈ Ei but is zero otherwise. The action of its transpose extends by zero a function defined on Ei to one defined on B. Corresponding to this partition of B, S can be written in the block form: SB = SE SEV ST EV SV . (5.9) The block SE can again be block partitioned, with most of the subblocks being zero. The diagonal blocks SEi of SE are the principal submatrices of S corresponding to Ei. Each SEi represents the coupling of nodes on interface Ei separating two neighboring subdomains. In defining the preconditioner, the action of S−1 Ei is needed. However, as noted before, in general SEi is a dense matrix which is also expensive to compute, and even if we had it, it would be expensive to compute its action (we would need to compute its inverse or a Cholesky factorization). Fortunately, many efficiently invertible approximations to SEi have been proposed in the literature (see Keyes and Gropp [135]) and we shall use these so-called interface preconditioners for SEi instead. We mention one specific preconditioner: MEi = αEi K1/2 where K is an nEi × nEi one dimensional Laplacian matrix, namely a tridiagonal matrix with 2’s down the main diagonal and −1’s down the two off-diagonals, and αEi is taken to be some average of the coefficient a(x, y) of (5.4) on the edge Ei. We note that since the eigen-decomposition of K is known and computable by the Fast Sine Transform, the action of MEi can be efficiently computed. With these notations, the Bramble-Pasciak-Schatz preconditioner is defined by its action on a vector vB defined on B as follows: K−1 BP SvB = RT HA−1 H RHvB + Ei RT Ei M−1 Ei REi vB. (5.10) Analogous to the additive Schwarz preconditioner, the computation of each term consists of the three steps of restriction-inversion-interpolation and is independent of the others and thus can be carried out in parallel. Bramble, Pasciak and Schatz [36] prove that the condition number of K−1 BP SSB is bounded by O(1 + log2 (H/h)). In practice, there is a slight growth in the number of iterations as h becomes small (i.e., as the fine grid is refined) or as H becomes large (i.e., as the coarse grid becomes coarser). The log2 (H/h) growth is due to the coupling of the unknowns on the edges incident on a com- mon vertex Vk, which is not accounted for in KBP S. Smith [190] proposed a vertex space modifica- tion to KBP S which explicitly accounts for this coupling and therefore eliminates the dependence on H and h. The idea is to introduce further subsets of B called vertex spaces X = k Xk with
- 89. 5.4. DOMAIN DECOMPOSITION METHODS 79 Xk consisting of a small set of vertices on the edges incident on the vertex Vk and adjacent to it. Note that X overlaps with E and VH. Let SXk be the principal submatrix of SB corresponding to Xk, and RXk , RT Xk be the corresponding restriction (pointwise) and extension (by zero) matri- ces. As before, SXk is dense and expensive to compute and factor/solve but efficiently invertable approximations (some using variants of the K1/2 operator defined before) have been developed in the literature (see Chan, Mathew and Shao [51]). We shall let MXk be such a preconditioner for SXk . Then Smith’s Vertex Space preconditioner is defined by: K−1 V SvB = RT HA−1 H RHvB + Ei RT Ei M−1 Ei REi vB + Xk RT Xk M−1 Xk RXk vB. (5.11) Smith [190] proved that the condition number of K−1 V SSB is bounded independent of H and h, provided there is sufficient overlap of Xk with B. 5.4.3 Further Remarks Multiplicative Schwarz Methods As mentioned before, the Additive Schwarz preconditioner can be viewed as an overlapping block Jacobi preconditioner. Analogously, one can define a multiplicative Schwarz preconditioner which corresponds to a symmetric block Gauss-Seidel version. That is, the solves on each subdomain are performed sequentially, using the most current iterates as boundary conditions from neighboring subdomains. Even without conjugate gradient acceleration, the multiplicative method can take many fewer iterations than the additive version. However, the multiplicative version is not as parallelizable, although the degree of parallelism can be increased by trading off the convergence rate through multi-coloring the subdomains. The theory can be found in Bramble, et al. [37]. Inexact Solves The exact solves involving A −1 i , A−1 i and A−1 H in Kas, KBP S, KV S can be replaced by inexact solves ˜A −1 i , ˜A−1 i and ˜A−1 H , which can be standard elliptic preconditioners themselves (e.g. multi- grid, ILU, SSOR, etc.). For the Schwarz methods, the modification is straightforward and the Inexact Solve Additive Schwarz Preconditioner is simply: ˜K−1 as v = RT H ˜A−1 H RHv + p i=1 RT i ˜A −1 i Riv. The Schur Complement methods require more changes to accommodate inexact solves. By replacing A−1 H by ˜A−1 H in the definitions of KBP S and KV S, we can easily obtain inexact pre- conditioners ˜KBP S and ˜KV S for SB. The main difficulty is, however, that the evaluation of the product SBvB requires exact subdomain solves in A−1 I . One way to get around this is to use an inner iteration using ˜Ai as a preconditioner for Ai in order to compute the action of A−1 I . An alternative is to perform the iteration on the larger system (5.5) and construct a preconditioner from the factorization in (5.6) by replacing the terms AI, SB by ˜AI, ˜SB respectively, where ˜SB can be either ˜KBP S or ˜KV S. Care must be taken to scale ˜AH and ˜Ai so that they are as close to AH and Ai as possible respectively — it is not sufficient that the condition number of ˜A−1 H AH and ˜A−1 i Ai be close to unity, because the scaling of the coupling matrix AIB may be wrong.
- 90. 80 CHAPTER 5. REMAINING TOPICS Nonsymmetric Problems The preconditioners given above extend naturally to nonsymmetric A’s (e.g., convection-diffusion problems), at least when the nonsymmetric part is not too large. The nice theoretical convergence rates can be retained provided that the coarse grid size H is chosen small enough (depending on the size of the nonsymmetric part of A) (see Cai and Widlund [43]). Practical implementations (especially for parallelism) of nonsymmetric domain decomposition methods are discussed in [136, 137]. Choice of Coarse Grid Size H Given h, it has been observed empirically (see Gropp and Keyes [110]) that there often exists an optimal value of H which minimizes the total computational time for solving the problem. A small H provides a better, but more expensive, coarse grid approximation, and requires solving more, but smaller, subdomain solves. A large H has the opposite effect. For model problems, the optimal H can be determined for both sequential and parallel implementations (see Chan and Shao [52]). In practice, it may pay to determine a near optimal value of H empirically if the preconditioner is to be re-used many times. However, there may also be geometric constraints on the range of values that H can take. 5.5 Multigrid Methods Simple iterative methods (such as the Jacobi method) tend to damp out high frequency components of the error fastest (see §2.2.1). This has led people to develop methods based on the following heuristic: 1. Perform some steps of a basic method in order to smooth out the error. 2. Restrict the current state of the problem to a subset of the grid points, the so-called “coarse grid”, and solve the resulting projected problem. 3. Interpolate the coarse grid solution back to the original grid, and perform a number of steps of the basic method again. Steps 1 and 3 are called “pre-smoothing” and “post-smoothing” respectively; by applying this method recursively to step 2 it becomes a true “multigrid” method. Usually the generation of subsequently coarser grids is halted at a point where the number of variables becomes small enough that direct solution of the linear system is feasible. The method outlined above is said to be a “V-cycle” method, since it descends through a se- quence of subsequently coarser grids, and then ascends this sequence in reverse order. A “W-cycle” method results from visiting the coarse grid twice, with possibly some smoothing steps in between. An analysis of multigrid methods is relatively straightforward in the case of simple differential operators such as the Poisson operator on tensor product grids. In that case, each next coarse grid is taken to have the double grid spacing of the previous grid. In two dimensions, a coarse grid will have one quarter of the number of points of the corresponding fine grid. Since the coarse grid is again a tensor product grid, a Fourier analysis (see for instance Briggs [42]) can be used. For the more general case of self-adjoint elliptic operators on arbitrary domains a more sophisticated analysis is needed (see Hackbusch [116], McCormick [147]). Many multigrid methods can be shown to have an (almost) optimal number of operations, that is, the work involved is proportional to the number of variables. From the above description it is clear that iterative methods play a role in multigrid theory as smoothers (see Kettler [132]). Conversely, multigrid-like methods can be used as preconditioners
- 91. 5.6. ROW PROJECTION METHODS 81 in iterative methods. The basic idea here is to partition the matrix on a given grid to a 2×2 structure A(i) = A (i) 1,1 A (i) 1,2 A (i) 2,1 A (i) 2,2 with the variables in the second block row corresponding to the coarse grid nodes. The matrix on the next grid is then an incomplete version of the Schur complement A(i+1) ≈ S(i) = A (i) 2,2 − A (i) 2,1A (i)−1 1,1 A (i) 1,2. The coarse grid is typically formed based on a red-black or cyclic reduction ordering; see for instance Rodrigue and Wolitzer [179], and Elman [92]. Some multigrid preconditioners try to obtain optimality results similar to those for the full multigrid method. Here we will merely supply some pointers to the literature: Axelsson and Eijk- hout [16], Axelsson and Vassilevski [23, 22], Braess [35], Maitre and Musy [144], McCormick and Thomas [148], Yserentant [216] and Wesseling [213]. 5.6 Row Projection Methods Most iterative methods depend on spectral properties of the coefficient matrix, for instance some require the eigenvalues to be in the right half plane. A class of methods without this limitation is that of row projection methods (see Bj¨orck and Elfving [34], and Bramley and Sameh [38]). They are based on a block row partitioning of the coefficient matrix AT = [A1, . . . , Am] and iterative application of orthogonal projectors Pix = Ai(AT i Ai)−1 AT i x. These methods have good parallel properties and seem to be robust in handling nonsymmetric and indefinite problems. Row projection methods can be used as preconditioners in the conjugate gradient method. In that case, there is a theoretical connection with the conjugate gradient method on the normal equa- tions (see §2.3.3).
- 92. 82 CHAPTER 5. REMAINING TOPICS
- 93. Appendix A Obtaining the Software A large body of numerical software is freely available 24 hours a day via an electronic service called Netlib. In addition to the template material, there are dozens of other libraries, technical reports on various parallel computers and software, test data, facilities to automatically translate FORTRAN programs to C, bibliographies, names and addresses of scientists and mathematicians, and so on. One can communicate with Netlib in one of a number of ways: by email, (much more easily) via an X-window interface called XNetlib, or through anonymous ftp (netlib2.cs.utk.edu). To get started using netlib, one sends a message of the form send index to [email protected]. A description of the entire library should be sent to you within minutes (providing all the interven- ing networks as well as the netlib server are up). FORTRAN and C versions of the templates for each method described in this book are available from Netlib. A user sends a request by electronic mail as follows: mail [email protected] On the subject line or in the body, single or multiple requests (one per line) may be made as follows: send index from linalg send sftemplates.shar from linalg The first request results in a return e-mail message containing the index from the library linalg, along with brief descriptions of its contents. The second request results in a return e-mail message consisting of a shar file containing the single precision FORTRAN routines and a README file. The versions of the templates that are available are listed in the table below: shar filename contents sctemplates.shar Single precision C routines dctemplates.shar Double precision C routines sftemplates.shar Single Precision Fortran 77 routines dftemplates.shar Double Precision Fortran 77 routines mltemplates.shar MATLAB routines Save the mail message to a file called templates.shar. Edit the mail header from this file and delete the lines down to and including << Cut Here >>. In the directory containing the shar file, type 83
- 94. 84 APPENDIX A. OBTAINING THE SOFTWARE sh templates.shar No subdirectory will be created. The unpacked files will stay in the current directory. Each method is written as a separate subroutine in its own file, named after the method (e.g., CG.f, BiCGSTAB.f, GMRES.f). The argument parameters are the same for each, with the exception of the required matrix-vector products and preconditioner solvers (some require the transpose matrix). Also, the amount of workspace needed varies. The details are documented in each routine. Note that the vector-vector operations are accomplished using the BLAS [143] (many manu- facturers have assembly coded these kernels for maximum performance), although a mask file is provided to link to user defined routines. The README file gives more details, along with instructions for a test routine.
- 95. Appendix B Overview of the BLAS The BLAS give us a standardized set of basic codes for performing operations on vectors and ma- trices. BLAS take advantage of the Fortran storage structure and the structure of the mathematical system wherever possible. Additionally, many computers have the BLAS library optimized to their system. Here we use five routines: 1. SCOPY: copies a vector onto another vector 2. SAXPY: adds vector x (multiplied by a scalar) to vector y 3. SGEMV: general matrix vector product 4. STRMV: matrix vector product when the matrix is triangular 5. STRSV: solves Tx = b for triangular matrix T The prefix “S” denotes single precision. This prefix may be changed to “D”, “C”, or “Z”, giving the routine double, complex, or double complex precision. (Of course, the declarations would also have to be changed.) It is important to note that putting double precision into single variables works, but single into double will cause errors. If we define ai,j = a(i,j) and xi = x(i), we can see what the code is doing: • ALPHA = SDOT( N, X, 1, Y, 1 ) computes the inner product of two vectors x and y, putting the result in scalar α. The corresponding Fortran segment is ALPHA = 0.0 DO I = 1, N ALPHA = ALPHA + X(I)*Y(I) ENDDO • CALL SAXPY( N, ALPHA, X, 1, Y ) multiplies a vector x of length n by the scalar α, then adds the result to the vector y, putting the result in y. The corresponding Fortran segment is DO I = 1, N Y(I) = ALPHA*X(I) + Y(I) ENDDO 85
- 96. 86 APPENDIX B. OVERVIEW OF THE BLAS • CALL SGEMV( ’N’, M, N, ONE, A, LDA, X, 1, ONE, B, 1 ) computes the matrix-vector product plus vector Ax + b, putting the resulting vector in b. The corresponding Fortran segment: DO J = 1, N DO I = 1, M B(I) = A(I,J)*X(J) + B(I) ENDDO ENDDO This illustrates a feature of the BLAS that often requires close attention. For example, we will use this routine to compute the residual vector b − Aˆx, where ˆx is our current approxi- mation to the solution x (merely change the fourth argument to -1.0E0). Vector b will be overwritten with the residual vector; thus, if we need it later, we will first copy it to temporary storage. • CALL STRMV( ’U’, ’N’, ’N’, N, A, LDA, X, 1 ) computes the matrix- vector product Ax, putting the resulting vector in x, for upper triangular matrix A. The corresponding Fortran segment is DO J = 1, N TEMP = X(J) DO I = 1, J X(I) = X(I) + TEMP*A(I,J) ENDDO ENDDO Note that the parameters in single quotes are for descriptions such as ’U’ for ‘UPPER TRI- ANGULAR’, ’N’ for ‘No Transpose’. This feature will be used extensively, resulting in storage savings (among other advantages). The variable LDA is critical for addressing the array correctly. LDA is the leading dimension of the two-dimensional array A, that is, LDA is the declared (or allocated) number of rows of the two-dimensional array A.
- 97. Appendix C Glossary Adaptive methods Iterative methods that collect information about the coefficient matrix during the iteration process, and use this to speed up convergence. Backward error The size of perturbations δA of the coefficient matrix and δb of the right hand side of a linear system Ax = b, such that the computed iterate x(i) is the solution of (A + δA)x(i) = b + δb. Band matrix A matrix A for which there are nonnegative constants p, q such that ai,j = 0 if j < i − p or j > i + q. The two constants p, q are called the left and right halfbandwidth respectively. Black box A piece of software that can be used without knowledge of its inner workings; the user supplies the input, and the output is assumed to be correct. BLAS Basic Linear Algebra Subprograms; a set of commonly occurring vector and matrix oper- ations for dense linear algebra. Optimized (usually assembly coded) implementations of the BLAS exist for various computers; these will give a higher performance than implementation in high level programming languages. Block factorization See: Block matrix operations. Block matrix operations Matrix operations expressed in terms of submatrices. Breakdown The occurrence of a zero divisor in an iterative method. Choleski decomposition Expressing a symmetric matrix A as a product of a lower triangular ma- trix L and its transpose LT , that is, A = LLT . Condition number See: Spectral condition number. Convergence The fact whether or not, or the rate at which, an iterative method approaches the solution of a linear system. Convergence can be • Linear: some measure of the distance to the solution decreases by a constant factor in each iteration. • Superlinear: the measure of the error decreases by a growing factor. • Smooth: the measure of the error decreases in all or most iterations, though not neces- sarily by the same factor. • Irregular: the measure of the error decreases in some iterations and increases in others. This observation unfortunately does not imply anything about the ultimate convergence of the method. 87
- 98. 88 APPENDIX C. GLOSSARY • Stalled: the measure of the error stays more or less constant during a number of iter- ations. As above, this does not imply anything about the ultimate convergence of the method. Dense matrix Matrix for which the number of zero elements is too small to warrant specialized algorithms to exploit these zeros. Diagonally dominant matrix See: Matrix properties Direct method An algorithm that produces the solution to a system of linear equations in a number of operations that is determined a priori by the size of the system. In exact arithmetic, a direct method yields the true solution to the system. See: Iterative method. Distributed memory See: Parallel computer. Divergence An iterative method is said to diverge if it does not converge in a reasonable number of iterations, or if some measure of the error grows unacceptably. However, growth of the error as such is no sign of divergence: a method with irregular convergence behavior may ultimately converge, even though the error grows during some iterations. Domain decomposition method Solution method for linear systems based on a partitioning of the physical domain of the differential equation. Domain decomposition methods typically involve (repeated) independent system solution on the subdomains, and some way of com- bining data from the subdomains on the separator part of the domain. Field of values Given a matrix A, the field of values is the set {xT Ax : xT x = 1}. For symmetric matrices this is the range [λmin(A), λmax(A)]. Fill A position that is zero in the original matrix A but not in an exact factorization of A. In an incomplete factorization, some fill elements are discarded. Forward error The difference between a computed iterate and the true solution of a linear system, measured in some vector norm. Halfbandwidth See: Band matrix. Ill-conditioned system A linear system for which the coefficient matrix has a large condition num- ber. Since in many applications the condition number is proportional to (some power of) the number of unknowns, this should be understood as the constant of proportionality being large. Incomplete factorization A factorization obtained by discarding certain elements during the fac- torization process (‘modified’ and ‘relaxed’ incomplete factorization compensate in some way for discarded elements). Thus an incomplete LU factorization of a matrix A will in general satisfy A = LU; however, one hopes that the factorization LU will be close enough to A to function as a preconditioner in an iterative method. Iterate Approximation to the solution of a linear system in any iteration of an iterative method. Iterative method An algorithm that produces a sequence of approximations to the solution of a linear system of equations; the length of the sequence is not given a priori by the size of the system. Usually, the longer one iterates, the closer one is able to get to the true solution. See: Direct method. Krylov sequence For a given matrix A and vector x, the sequence of vectors {Ai x}i≥0, or a finite initial part of this sequence.
- 99. 89 Krylov subspace The subspace spanned by a Krylov sequence. LAPACK A mathematical library of linear algebra routine for dense systems solution and eigen- value calculations. Lower triangular matrix Matrix A for which ai,j = 0 if j > i. LQ factorization A way of writing a matrix A as a product of a lower triangular matrix L and a unitary matrix Q, that is, A = LQ. LU factorization / LU decomposition Expressing a matrix A as a product of a lower triangular matrix L and an upper triangular matrix U, that is, A = LU. M-Matrix See: Matrix properties. Matrix norms See: Norms. Matrix properties We call a square matrix A Symmetric if ai,j = aj,i for all i, j. Positive definite if it satisfies xT Ax > 0 for all nonzero vectors x. Diagonally dominant if ai,i > j=i |ai,j|; the excess amount mini{ai,i − j=i |ai,j|} is called the diagonal dominance of the matrix. An M-matrix if ai,j ≤ 0 for i = j, and it is nonsingular with (A−1 )i,j ≥ 0 for all i, j. Message passing See: Parallel computer. Multigrid method Solution method for linear systems based on restricting and extrapolating so- lutions between a series of nested grids. Modified incomplete factorization See: Incomplete factorization. Mutually consistent norms See: Norms. Natural ordering See: Ordering of unknowns. Nonstationary iterative method Iterative method that has iteration-dependent coefficients. Normal equations For a nonsymmetric or indefinite (but nonsingular) system of equations Ax = b, either of the related symmetric systems (AT Ax = AT b) and (AAT y = b; x = AT y). For complex A, AT is replaced with AH in the above expressions. Norms A function f : Rn → R is called a vector norm if • f(x) ≥ 0 for all x, and f(x) = 0 only if x = 0. • f(αx) = |α|f(x) for all α, x. • f(x + y) ≤ f(x) + f(y) for all x, y. The same properties hold for matrix norms. A matrix norm and a vector norm (both de- noted · ) are called a mutually consistent pair if for all matrices A and vectors x Ax ≤ A x . Ordering of unknowns For linear systems derived from a partial differential equation, each un- known corresponds to a node in the discretization mesh. Different orderings of the unknowns correspond to permutations of the coefficient matrix. The convergence speed of iterative methods may depend on the ordering used, and often the parallel efficiency of a method on a parallel computer is strongly dependent on the ordering used. Some common orderings for rectangular domains are:
- 100. 90 APPENDIX C. GLOSSARY • The natural ordering; this is the consecutive numbering by rows and columns. • The red/black ordering; this is the numbering where all nodes with coordinates (i, j) for which i + j is odd are numbered before those for which i + j is even. • The ordering by diagonals; this is the ordering where nodes are grouped in levels for which i + j is constant. All nodes in one level are numbered before the nodes in the next level. For matrices from problems on less regular domains, some common orderings are: • The Cuthill-McKee ordering; this starts from one point, then numbers its neighbors, and continues numbering points that are neighbors of already numbered points. The Re- verse Cuthill-McKee ordering then reverses the numbering; this may reduce the amount of fill in a factorization of the matrix. • The Minimum Degree ordering; this orders the matrix rows by increasing numbers of nonzeros. Parallel computer Computer with multiple independent processing units. If the processors have immediate access to the same memory, the memory is said to be shared; if processors have private memory that is not immediately visible to other processors, the memory is said to be distributed. In that case, processors communicate by message-passing. Pipelining See: Vector computer. Positive definite matrix See: Matrix properties. Preconditioner An auxiliary matrix in an iterative method that approximates in some sense the coefficient matrix or its inverse. The preconditioner, or preconditioning matrix, is applied in every step of the iterative method. Red/black ordering See: Ordering of unknowns. Reduced system Linear system obtained by eliminating certain variables from another linear sys- tem. Although the number of variables is smaller than for the original system, the matrix of a reduced system generally has more nonzero entries. If the original matrix was symmetric and positive definite, then the reduced system has a smaller condition number. Relaxed incomplete factorization See: Incomplete factorization. Residual If an iterative method is employed to solve for x in a linear system Ax = b, then the residual corresponding to a vector y is Ay − b. Search direction Vector that is used to update an iterate. Shared memory See: Parallel computer. Simultaneous displacements, method of Jacobi method. Sparse matrix Matrix for which the number of zero elements is large enough that algorithms avoiding operations on zero elements pay off. Matrices derived from partial differential equations typically have a number of nonzero elements that is proportional to the matrix size, while the total number of matrix elements is the square of the matrix size. Spectral condition number The product A 2 · A−1 2 = λ 1/2 max(AT A) λ 1/2 min(AT A) ,
- 101. C.1. NOTATION 91 where λmax and λmin denote the largest and smallest eigenvalues, respectively. For linear systems derived from partial differential equations in 2D, the condition number is propor- tional to the number of unknowns. Spectral radius The spectral radius of a matrix A is max{|λ(A)|}. Spectrum The set of all eigenvalues of a matrix. Stationary iterative method Iterative method that performs in each iteration the same operations on the current iteration vectors. Stopping criterion Since an iterative method computes successive approximations to the solution of a linear system, a practical test is needed to determine when to stop the iteration. Ideally this test would measure the distance of the last iterate to the true solution, but this is not possible. Instead, various other metrics are used, typically involving the residual. Storage scheme The way elements of a matrix are stored in the memory of a computer. For dense matrices, this can be the decision to store rows or columns consecutively. For sparse matrices, common storage schemes avoid storing zero elements; as a result they involve indices, stored as integer data, that indicate where the stored elements fit into the global matrix. Successive displacements, method of Gauss-Seidel method. Symmetric matrix See: Matrix properties. Template Description of an algorithm, abstracting away from implementational details. Tune Adapt software for a specific application and computing environment in order to obtain better performance in that case only. itemize Upper triangular matrix Matrix A for which ai,j = 0 if j < i. Vector computer Computer that is able to process consecutive identical operations (typically ad- ditions or multiplications) several times faster than intermixed operations of different types. Processing identical operations this way is called ‘pipelining’ the operations. Vector norms See: Norms. C.1 Notation In this section, we present some of the notation we use throughout the book. We have tried to use standard notation that would be found in any current publication on the subjects covered. Throughout, we follow several conventions: • Matrices are denoted by capital letters. • Vectors are denoted by lowercase letters. • Lowercase greek letters usually denote scalars, for instance • Matrix elements are denoted by doubly indexed lowercase letter, however • Matrix subblocks are denoted by doubly indexed uppercase letters.
- 102. 92 APPENDIX C. GLOSSARY We define matrix A of dimension m × n and block dimension m × n as follows: A = a1,1 · · · a1,n ... ... am,1 · · · am,n (ai,j ∈ R) A1,1 · · · A1,n ... ... Am ,1 · · · Am ,n (Ai,j ∈ R × ). We define vector x of dimension n as follows: x = x1 ... xn xi ∈ R. Other notation is as follows: • In×n (or simply I if the size is clear from the context) denotes the identity matrix. • A = diag(ai,i) denotes that matrix A has elements ai,i on its diagonal, and zeros everywhere else. • x (k) i denotes the ith element of vector x during the kth iteration
- 103. Bibliography [1] J. AARDEN AND K.-E. KARLSSON, Preconditioned CG-type methods for solving the cou- pled systems of fundamental semiconductor equations, BIT, 29 (1989), pp. 916–937. [2] L. ADAMS AND H. JORDAN, Is SOR color-blind?, SIAM J. Sci. Statist. Comput., 7 (1986), pp. 490–506. [3] E. ANDERSON, ET. AL., LAPACK Users Guide, SIAM, Philadelphia, 1992. [4] J. APPLEYARD AND I. CHESHIRE, Nested factorization, in Reservoir Simulation Sympo- sium of the SPE, 1983. Paper 12264. [5] M. ARIOLI, J. DEMMEL, AND I. DUFF, Solving sparse linear systems with sparse backward error, SIAM J. Matrix Anal. Appl., 10 (1989), pp. 165–190. [6] W. ARNOLDI, The principle of minimized iterations in the solution of the matrix eigenvalue problem, Quart. Appl. Math., 9 (1951), pp. 17–29. [7] S. ASHBY, CHEBYCODE: A Fortran implementation of Manteuffel’s adaptive Chebyshev algorithm, Tech. Rep. UIUCDCS-R-85-1203, University of Illinois, 1985. [8] S. ASHBY, T. MANTEUFFEL, AND J. OTTO, A comparison of adaptive Chebyshev and least squares polynomial preconditioning for Hermitian positive definite linear systems, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 1–29. [9] S. ASHBY, T. MANTEUFFEL, AND P. SAYLOR, Adaptive polynomial preconditioning for Hermitian indefinite linear systems, BIT, 29 (1989), pp. 583–609. [10] S. F. ASHBY, T. A. MANTEUFFEL, AND P. E. SAYLOR, A taxonomy for conjugate gradient methods, SIAM J. Numer. Anal., 27 (1990), pp. 1542–1568. [11] C. ASHCRAFT AND R. GRIMES, On vectorizing incomplete factorizations and SSOR pre- conditioners, SIAM J. Sci. Statist. Comput., 9 (1988), pp. 122–151. [12] O. AXELSSON, Incomplete block matrix factorization preconditioning methods. The ulti- mate answer?, J. Comput. Appl. Math., 12&13 (1985), pp. 3–18. [13] , A general incomplete block-matrix factorization method, Linear Algebra Appl., 74 (1986), pp. 179–190. [14] O. AXELSSON AND A. BARKER, Finite element solution of boundary value problems. The- ory and computation, Academic Press, Orlando, Fl., 1984. [15] O. AXELSSON AND V. EIJKHOUT, Vectorizable preconditioners for elliptic difference equa- tions in three space dimensions, J. Comput. Appl. Math., 27 (1989), pp. 299–321. 93
- 104. 94 BIBLIOGRAPHY [16] , The nested recursive two-level factorization method for nine-point difference matri- ces, SIAM J. Sci. Statist. Comput., 12 (1991), pp. 1373–1400. [17] O. AXELSSON AND I. GUSTAFSSON, Iterative solution for the solution of the Navier equa- tions of elasticity, Comput. Methods Appl. Mech. Engrg., 15 (1978), pp. 241–258. [18] O. AXELSSON AND G. LINDSKOG, On the eigenvalue distribution of a class of precondi- tioning matrices, Numer. Math., 48 (1986), pp. 479–498. [19] , On the rate of convergence of the preconditioned conjugate gradient method, Numer. Math., 48 (1986), pp. 499–523. [20] O. AXELSSON AND N. MUNKSGAARD, Analysis of incomplete factorizations with fixed storage allocation, in Preconditioning Methods – Theory and Applications, D. Evans, ed., Gordon and Breach, New York, 1983, pp. 265–293. [21] O. AXELSSON AND B. POLMAN, On approximate factorization methods for block-matrices suitable for vector and parallel processors, Linear Algebra Appl., 77 (1986), pp. 3–26. [22] O. AXELSSON AND P. VASSILEVSKI, Algebraic multilevel preconditioning methods, I, Nu- mer. Math., 56 (1989), pp. 157–177. [23] , Algebraic multilevel preconditioning methods, II, SIAM J. Numer. Anal., 57 (1990), pp. 1569–1590. [24] O. AXELSSON AND P. S. VASSILEVSKI, A black box generalized conjugate gradient solver with inner iterations and variable-step preconditioning, SIAM J. Matrix Anal. Appl., 12 (1991), pp. 625–644. [25] R. BANK, Marching algorithms for elliptic boundary value problems; II: The variable co- efficient case, SIAM J. Numer. Anal., 14 (1977), pp. 950–970. [26] R. BANK, T. CHAN, W. COUGHRAN JR., AND R. SMITH, The Alternate-Block- Factorization procedure for systems of partial differential equations, BIT, 29 (1989), pp. 938–954. [27] R. BANK AND D. ROSE, Marching algorithms for elliptic boundary value problems. I: The constant coefficient case, SIAM J. Numer. Anal., 14 (1977), pp. 792–829. [28] R. E. BANK AND T. F. CHAN, A composite step bi-conjugate gradient algorithm for non- symmetric linear systems, Tech. Rep. CAM 92-, UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1992. [29] R. E. BANK AND T. F. CHAN, An analysis of the composite step Biconjugate gradient method, Numerische Mathematik, 66 (1993), pp. 295–319. [30] G. BAUDET, Asynchronous iterative methods for multiprocessors, J. Assoc. Comput. Mach., 25 (1978), pp. 226–244. [31] R. BEAUWENS, On Axelsson’s perturbations, Linear Algebra Appl., 68 (1985), pp. 221– 242. [32] , Approximate factorizations with S/P consistently ordered M-factors, BIT, 29 (1989), pp. 658–681. [33] R. BEAUWENS AND L. QUENON, Existence criteria for partial matrix factorizations in iterative methods, SIAM J. Numer. Anal., 13 (1976), pp. 615–643.
- 105. BIBLIOGRAPHY 95 [34] A. BJ ¨ORCK AND T. ELFVING, Accelerated projection methods for computing pseudo- inverse solutions of systems of linear equations, BIT, 19 (1979), pp. 145–163. [35] D. BRAESS, The contraction number of a multigrid method for solving the Poisson equation, Numer. Math., 37 (1981), pp. 387–404. [36] J. H. BRAMBLE, J. E. PASCIAK, AND A. H. SCHATZ, The construction of preconditioners for elliptic problems by substructuring, I, Mathematics of Computation, 47 (1986), pp. 103– 134. [37] J. H. BRAMBLE, J. E. PASCIAK, J. WANG, AND J. XU, Convergence estimates for product iterative methods with applications to domain decompositions and multigrid, Math. Comp., to appear. [38] R. BRAMLEY AND A. SAMEH, Row projection methods for large nonsymmetric linear sys- tems, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 168–193. [39] C. BREZINSKI AND H. SADOK, Avoiding breakdown in the CGS algorithm, Numer. Alg., 1 (1991), pp. 199–206. [40] C. BREZINSKI, M. ZAGLIA, AND H. SADOK, Avoiding breakdown and near breakdown in Lanczos type algorithms, Numer. Alg., 1 (1991), pp. 261–284. [41] , A breakdown free Lanczos type algorithm for solving linear systems, Numer. Math., 63 (1992), pp. 29–38. [42] W. BRIGGS, A Multigrid Tutorial, SIAM, Philadelphia, 1977. [43] X.-C. CAI AND O. WIDLUND, Multiplicative Schwarz algorithms for some nonsymmetric and indefinite problems, SIAM J. Numer. Anal., 30 (1993), pp. 936–952. [44] T. CHAN, Fourier analysis of relaxed incomplete factorization preconditioners, SIAM J. Sci. Statist. Comput., 12 (1991), pp. 668–680. [45] T. CHAN, L. DE PILLIS, AND H. VAN DER VORST, A transpose-free squared Lanczos algorithm and application to solving nonsymmetric linear systems, Tech. Rep. CAM 91-17, UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1991. [46] T. CHAN, E. GALLOPOULOS, V. SIMONCINI, T. SZETO, AND C. TONG, A quasi-minimal residual variant of the Bi-CGSTAB algorithm for nonsymmetric systems, Tech. Rep. CAM 92-26, UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1992. SIAM J. Sci. Comput., to appear. [47] T. CHAN, R. GLOWINSKI, , J. P´ERIAUX, AND O. WIDLUND, eds., Domain Decomposition Methods, Philadelphia, 1989, SIAM. Proceedings of the Second International Symposium on Domain Decomposition Methods, Los Angeles, CA, January 14 - 16, 1988. [48] , eds., Domain Decomposition Methods, Philadelphia, 1990, SIAM. Proceedings of the Third International Symposium on Domain Decomposition Methods, Houston, TX, 1989. [49] , eds., Domain Decomposition Methods, SIAM, Philadelphia, 1991. Proceedings of the Fourth International Symposium on Domain Decomposition Methods, Moscow, USSR, 1990. [50] T. CHAN AND C.-C. J. KUO, Two-color Fourier analysis of iterative algorithms for elliptic problems with red/black ordering, SIAM J. Sci. Statist. Comput., 11 (1990), pp. 767–793.
- 106. 96 BIBLIOGRAPHY [51] T. F. CHAN, T. P. MATHEW, AND J. P. SHAO, Efficient variants of the vertex space domain decomposition algorithm, Tech. Rep. CAM 92-07, UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1992. SIAM J. Sci. Comput., to appear. [52] T. F. CHAN AND J. SHAO, On the choice of coarse grid size in domain decomposition methods, tech. rep., UCLA, Dept. of Math., Los Angeles, CA 90024-1555, 1993. to appear. [53] D. CHAZAN AND W. MIRANKER, Chaotic relaxation, Linear Algebra Appl., 2 (1969), pp. 199–222. [54] A. CHRONOPOULOS AND C. GEAR, s-step iterative methods for symmetric linear systems, J. Comput. Appl. Math., 25 (1989), pp. 153–168. [55] P. CONCUS AND G. GOLUB, A generalized conjugate gradient method for nonsymmet- ric systems of linear equations, in Computer methods in Applied Sciences and Engineering, Second International Symposium, Dec 15–19, 1975; Lecture Notes in Economics and Math- ematical Systems, Vol. 134, Berlin, New York, 1976, Springer-Verlag. [56] P. CONCUS, G. GOLUB, AND G. MEURANT, Block preconditioning for the conjugate gra- dient method, SIAM J. Sci. Statist. Comput., 6 (1985), pp. 220–252. [57] P. CONCUS, G. GOLUB, AND D. O’LEARY, A generalized conjugate gradient method for the numerical solution of elliptic partial differential equations, in Sparse Matrix Computa- tions, J. Bunch and D. Rose, eds., Academic Press, New York, 1976, pp. 309–332. [58] P. CONCUS AND G. H. GOLUB, Use of fast direct methods for the efficient numerical solu- tion of nonseparable elliptic equations, SIAM J. Numer. Anal., 10 (1973), pp. 1103–1120. [59] E. CUTHILL AND J. MCKEE, Reducing the bandwidth of sparse symmetric matrices, in ACM Proceedings of the 24th National Conference, 1969. [60] E. D’AZEVEDO, V. EIJKHOUT, AND C. ROMINE, LAPACK working note 56: Reducing communication costs in the conjugate gradient algorithm on distributed memory multipro- cessor, tech. report, Computer Science Department, University of Tennessee, Knoxville, TN, 1993. [61] E. D’AZEVEDO AND C. ROMINE, Reducing communication costs in the conjugate gradient algorithm on distributed memory multiprocessors, Tech. Rep. ORNL/TM-12192, Oak Ridge National Lab, Oak Ridge, TN, 1992. [62] E. DE STURLER, A parallel restructured version of GMRES(m), Tech. Rep. 91-85, Delft University of Technology, Delft, The Netherlands, 1991. [63] E. DE STURLER AND D. R. FOKKEMA, Nested Krylov methods and preserving the orthog- onality, Tech. Rep. Preprint 796, Utrecht University, Utrecht, The Netherlands, 1993. [64] S. DEMKO, W. MOSS, AND P. SMITH, Decay rates for inverses of band matrices, Mathe- matics of Computation, 43 (1984), pp. 491–499. [65] J. DEMMEL, The condition number of equivalence transformations that block diagonalize matrix pencils, SIAM J. Numer. Anal., 20 (1983), pp. 599–610. [66] J. DEMMEL, M. HEATH, AND H. VAN DER VORST, Parallel linear algebra, in Acta Nu- merica, Vol. 2, Cambridge Press, New York, 1993. [67] S. DOI, On parallelism and convergence of incomplete LU factorizations, Appl. Numer. Math., 7 (1991), pp. 417–436.
- 107. BIBLIOGRAPHY 97 [68] J. DONGARRA, J. DUCROZ, I. DUFF, AND S. HAMMARLING, A set of level 3 Basic Linear Algebra Subprograms, ACM Trans. Math. Soft., 16 (1990), pp. 1–17. [69] J. DONGARRA, J. DUCROZ, S. HAMMARLING, AND R. HANSON, An extended set of FORTRAN Basic Linear Algebra Subprograms, ACM Trans. Math. Soft., 14 (1988), pp. 1– 32. [70] J. DONGARRA, I. DUFF, D. SORENSEN, AND H. VAN DER VORST, Solving Linear Systems on Vector and Shared Memory Computers, SIAM, Philadelphia, PA, 1991. [71] J. DONGARRA AND E. GROSSE, Distribution of mathematical software via electronic mail, Comm. ACM, 30 (1987), pp. 403–407. [72] J. DONGARRA, C. MOLER, J. BUNCH, AND G. STEWART, LINPACK Users’ Guide, SIAM, Philadelphia, 1979. [73] J. DONGARRA AND H. VAN DER VORST, Performance of various computers using stan- dard sparse linear equations solving techniques, in Computer Benchmarks, J. Dongarra and W. Gentzsch, eds., Elsevier Science Publishers B.V., New York, 1993, pp. 177–188. [74] F. DORR, The direct solution of the discrete Poisson equation on a rectangle, SIAM Rev., 12 (1970), pp. 248–263. [75] M. DRYJA AND O. B. WIDLUND, Towards a unified theory of domain decomposition algo- rithms for elliptic problems, Tech. Rep. 486, also Ultracomputer Note 167, Department of Computer Science, Courant Institute, 1989. [76] D. DUBOIS, A. GREENBAUM, AND G. RODRIGUE, Approximating the inverse of a matrix for use in iterative algorithms on vector processors, Computing, 22 (1979), pp. 257–268. [77] I. DUFF, R. GRIMES, AND J. LEWIS, Sparse matrix test problems, ACM Trans. Math. Soft., 15 (1989), pp. 1–14. [78] I. DUFF AND G. MEURANT, The effect of ordering on preconditioned conjugate gradients, BIT, 29 (1989), pp. 635–657. [79] I. S. DUFF, A. M. ERISMAN, AND J.K.REID, Direct methods for sparse matrices, Oxford University Press, London, 1986. [80] T. DUPONT, R. KENDALL, AND H. RACHFORD, An approximate factorization proce- dure for solving self-adjoint elliptic difference equations, SIAM J. Numer. Anal., 5 (1968), pp. 559–573. [81] E. D’YAKONOV, The method of variable directions in solving systems of finite difference equations, Soviet Math. Dokl., 2 (1961), pp. 577–580. TOM 138, 271–274. [82] L. EHRLICH, An Ad-Hoc SOR method, J. Comput. Phys., 43 (1981), pp. 31–45. [83] M. EIERMANN AND R. VARGA, Is the optimal ω best for the SOR iteration method?, Linear Algebra Appl., 182 (1993), pp. 257–277. [84] V. EIJKHOUT, Analysis of parallel incomplete point factorizations, Linear Algebra Appl., 154–156 (1991), pp. 723–740. [85] , Beware of unperturbed modified incomplete point factorizations, in Proceedings of the IMACS International Symposium on Iterative Methods in Linear Algebra, Brussels, Bel- gium, R. Beauwens and P. de Groen, eds., 1992.
- 108. 98 BIBLIOGRAPHY [86] , LAPACK working note 50: Distributed sparse data structures for linear algebra operations, Tech. Rep. CS 92-169, Computer Science Department, University of Tennessee, Knoxville, TN, 1992. [87] , LAPACK working note 51: Qualitative properties of the conjugate gradient and Lanc- zos methods in a matrix framework, Tech. Rep. CS 92-170, Computer Science Department, University of Tennessee, Knoxville, TN, 1992. [88] V. EIJKHOUT AND B. POLMAN, Decay rates of inverses of banded M-matrices that are near to Toeplitz matrices, Linear Algebra Appl., 109 (1988), pp. 247–277. [89] V. EIJKHOUT AND P. VASSILEVSKI, Positive definitess aspects of vectorizable precondi- tioners, Parallel Computing, 10 (1989), pp. 93–100. [90] S. EISENSTAT, Efficient implementation of a class of preconditioned conjugate gradient methods, SIAM J. Sci. Statist. Comput., 2 (1981), pp. 1–4. [91] R. ELKIN, Convergence theorems for Gauss-Seidel and other minimization algorithms, Tech. Rep. 68-59, Computer Science Center, University of Maryland, College Park, MD, Jan. 1968. [92] H. ELMAN, Approximate Schur complement preconditioners on serial and parallel comput- ers, SIAM J. Sci. Statist. Comput., 10 (1989), pp. 581–605. [93] H. ELMAN AND M. SCHULTZ, Preconditioning by fast direct methods for non self-adjoint nonseparable elliptic equations, SIAM J. Numer. Anal., 23 (1986), pp. 44–57. [94] L. ELSNER, A note on optimal block-scaling of matrices, Numer. Math., 44 (1984), pp. 127– 128. [95] V. FABER AND T. MANTEUFFEL, Necessary and sufficient conditions for the existence of a conjugate gradient method, SIAM J. Numer. Anal., 21 (1984), pp. 315–339. [96] G. FAIRWEATHER, A. GOURLAY, AND A. MITCHELL, Some high accuracy difference schemes with a splitting operator for equations of parabolic and elliptic type, Numer. Math., 10 (1967), pp. 56–66. [97] R. FLETCHER, Conjugate gradient methods for indefinite systems, in Numerical Analysis Dundee 1975, G. Watson, ed., Berlin, New York, 1976, Springer Verlag, pp. 73–89. [98] G. FORSYTHE AND E. STRAUSS, On best conditioned matrices, Proc. Amer. Math. Soc., 6 (1955), pp. 340–345. [99] R. FREUND, Conjugate gradient-type methods for linear systems with complex symmetric coefficient matrices, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 425–448. [100] R. FREUND, M. GUTKNECHT, AND N. NACHTIGAL, An implementation of the look-ahead Lanczos algorithm for non-Hermitian matrices, SIAM J. Sci. Comput., 14 (1993), pp. 137– 158. [101] R. FREUND AND N. NACHTIGAL, QMR: A quasi-minimal residual method for non- Hermitian linear systems, Numer. Math., 60 (1991), pp. 315–339. [102] , An implementation of the QMR method based on coupled two-term recurrences, Tech. Rep. 92.15, RIACS, NASA Ames, Ames, CA, 1992. [103] R. FREUND AND T. SZETO, A quasi-minimal residual squared algorithm for non-Hermitian linear systems, tech. rep., RIACS, NASA Ames, Ames, CA, 1991.
- 109. BIBLIOGRAPHY 99 [104] R. W. FREUND, A transpose-free quasi-minimum residual algorithm for non-Hermitian lin- ear systems, SIAM J. Sci. Comput., 14 (1993), pp. 470–482. [105] R. W. FREUND, G. H. GOLUB, AND N. M. NACHTIGAL, Iterative solution of linear sys- tems, Acta Numerica, (1992), pp. 57–100. [106] R. GLOWINSKI, G. H. GOLUB, G. A. MEURANT, AND J. P´ERIAUX, eds., Domain Decom- position Methods for Partial Differential Equations, SIAM, Philadelphia, 1988. Proceedings of the First International Symposium on Domain Decomposition Methods for Partial Differ- ential Equations, Paris, France, January 1987. [107] G. GOLUB AND D. O’LEARY, Some history of the conjugate gradient and Lanczos meth- ods, SIAM Rev., 31 (1989), pp. 50–102. [108] G. GOLUB AND C. VAN LOAN, Matrix Computations, second edition, The Johns Hopkins University Press, Baltimore, 1989. [109] A. GREENBAUM AND Z. STRAKOS, Predicting the behavior of finite precision Lanczos and conjugate gradient computations, SIAM J. Mat. Anal. Appl., 13 (1992), pp. 121–137. [110] W. D. GROPP AND D. E. KEYES, Domain decomposition with local mesh refinement, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 967–993. [111] I. GUSTAFSSON, A class of first-order factorization methods, BIT, 18 (1978), pp. 142–156. [112] M. H. GUTKNECHT, A completed theory of the unsymmetric Lanczos process and related algorithms, part II, Tech. Rep. 90-16, IPS Research Report, ETH Z¨urich, Switzerland, 1990. [113] , The unsymmetric Lanczos algorithms and their relations to P´ade approximation, con- tinued fractions and the QD algorithm, in Proceedings of the Copper Mountain Conference on Iterative Methods, 1990. [114] , Variants of Bi-CGSTAB for matrices with complex spectrum, Tech. Rep. 91-14, IPS ETH, Z¨urich, Switzerland, 1991. [115] , A completed theory of the unsymmetric Lanczos process and related algorithms, part I, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 594–639. [116] W. HACKBUSCH, Multi-Grid Methods and Applications, Springer-Verlag, Berlin, New York, 1985. [117] , Iterative L¨osung großer schwachbesetzter Gleichungssysteme, Teubner, Stuttgart, 1991. [118] A. HADJIDIMOS, On some high accuracy difference schemes for solving elliptic equations, Numer. Math., 13 (1969), pp. 396–403. [119] L. HAGEMAN AND D. YOUNG, Applied Iterative Methods, Academic Press, New York, 1981. [120] W. HAGER, Condition estimators, SIAM J. Sci. Statist. Comput., 5 (1984), pp. 311–316. [121] M. HESTENES AND E. STIEFEL, Methods of conjugate gradients for solving linear systems, J. Res. Nat. Bur. Stand., 49 (1952), pp. 409–436. [122] M. R. HESTENES, Conjugacy and gradients, in A History of Scientific Computing, Addison-Wesley, Reading, MA, 1990, pp. 167–179.
- 110. 100 BIBLIOGRAPHY [123] N. HIGHAM, Experience with a matrix norm estimator, SIAM J. Sci. Statist. Comput., 11 (1990), pp. 804–809. [124] K. JEA AND D. YOUNG, Generalized conjugate-gradient acceleration of nonsym- metriz- able iterative methods, Linear Algebra Appl., 34 (1980), pp. 159–194. [125] O. JOHNSON, C. MICCHELLI, AND G. PAUL, Polynomial preconditioning for conjugate gradient calculation, SIAM J. Numer. Anal., 20 (1983), pp. 362–376. [126] M. JONES AND P. PLASSMANN, Parallel solution of unstructed, sparse systems of linear equations, in Proceedings of the Sixth SIAM conference on Parallel Processing for Scien- tific Computing, R. Sincovec, D. Keyes, M. Leuze, L. Petzold, and D. Reed, eds., SIAM, Philadelphia, pp. 471–475. [127] , A parallel graph coloring heuristic, SIAM J. Sci. Statist. Comput., 14 (1993), pp. 654–669. [128] W. JOUBERT, Lanczos methods for the solution of nonsymmetric systems of linear equations, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 926–943. [129] W. KAHAN, Gauss-Seidel methods of solving large systems of linear equations, PhD thesis, University of Toronto, 1958. [130] S. KANIEL, Estimates for some computational techniques in linear algebra, Mathematics of Computation, 20 (1966), pp. 369–378. [131] D. KERSHAW, The incomplete Cholesky-conjugate gradient method for the iterative solu- tion of systems of linear equations, J. Comput. Phys., 26 (1978), pp. 43–65. [132] R. KETTLER, Analysis and comparison of relaxation schemes in robust multigrid and pre- conditioned conjugate gradient methods, in Multigrid Methods, Lecture Notes in Mathemat- ics 960, W. Hackbusch and U. Trottenberg, eds., Springer-Verlag, Berlin, New York, 1982, pp. 502–534. [133] , Linear multigrid methods in numerical reservoir simulation, PhD thesis, Delft Uni- versity of Technology, Delft, The Netherlands, 1987. [134] D. E. KEYES, T. F. CHAN, G. MEURANT, J. S. SCROGGS, AND R. G. VOIGT, eds., Domain Decomposition Methods For Partial Differential Equations, SIAM, Philadelphia, 1992. Proceedings of the Fifth International Symposium on Domain Decomposition Meth- ods, Norfolk, VA, 1991. [135] D. E. KEYES AND W. D. GROPP, A comparison of domain decomposition techniques for elliptic partial differential equations and their parallel implementation, SIAM J. Sci. Statist. Comput., 8 (1987), pp. s166 – s202. [136] , Domain decomposition for nonsymmetric systems of equations: Examples from com- putational fluid dynamics, in Domain Decomposition Methods, proceedings of the Sec- ond Internation Symposium, Los Angeles, California, January 14–16, 1988, T. F. Chan, R. Glowinski, J. Periaux, and O. B. Widlund, eds., Philadelphia, 1989, SIAM, pp. 373–384. [137] , Domain decomposition techniques for the parallel solution of nonsymmetric systems of elliptic boundary value problems, Applied Num. Math., 6 (1989/1990), pp. 281–301. [138] S. K. KIM AND A. T. CHRONOPOULOS, A class of Lanczos-like algorithms implemented on parallel computers, Parallel Comput., 17 (1991), pp. 763–778.
- 111. BIBLIOGRAPHY 101 [139] D. R. KINCAID, J. R. RESPESS, D. M. YOUNG, AND R. G. GRIMES, ITPACK 2C: A Fortran package for solving large sparse linear systems by adaptive accelerated iterative methods, ACM Trans. Math. Soft., 8 (1982), pp. 302–322. Algorithm 586. [140] L. Y. KOLOTILINA AND A. Y. YEREMIN, On a family of two-level preconditionings of the incomlete block factorization type, Sov. J. Numer. Anal. Math. Modelling, (1986), pp. 293– 320. [141] C. LANCZOS, An iteration method for the solution of the eigenvalue problem of linear dif- ferential and integral operators, J. Res. Nat. Bur. Stand., 45 (1950), pp. 255–282. [142] , Solution of systems of linear equations by minimized iterations, J. Res. Nat. Bur. Stand., 49 (1952), pp. 33–53. [143] C. LAWSON, R. HANSON, D. KINCAID, AND F. KROGH, Basic Linear Algebra Subpro- grams for FORTRAN usage, ACM Trans. Math. Soft., 5 (1979), pp. 308–325. [144] J. MAITRE AND F. MUSY, The contraction number of a class of two-level methods; an exact evaluation for some finite element subspaces and model problems, in Multigrid methods, Proceedings, K¨oln-Porz, 1981, W. Hackbusch and U. Trottenberg, eds., vol. 960 of Lecture Notes in Mathematics, 1982, pp. 535–544. [145] T. MANTEUFFEL, The Tchebychev iteration for nonsymmetric linear systems, Numer. Math., 28 (1977), pp. 307–327. [146] , An incomplete factorization technique for positive definite linear systems, Mathemat- ics of Computation, 34 (1980), pp. 473–497. [147] S. MCCORMICK, Multilevel Adaptive Methods for Partial Differential Equations, SIAM, Philadelphia, 1989. [148] S. MCCORMICK AND J. THOMAS, The Fast Adaptive Composite grid (FAC) method for elliptic equations, Mathematics of Computation, 46 (1986), pp. 439–456. [149] U. MEIER AND A. SAMEH, The behavior of conjugate gradient algorithms on a multivector processor with a hierarchical memory, J. Comput. Appl. Math., 24 (1988), pp. 13–32. [150] U. MEIER-YANG, Preconditioned conjugate gradient-like methods for nonsymmetric linear systems, tech. rep., CSRD, University of Illinois, Urbana, IL, April 1992. [151] J. MEIJERINK AND H. VAN DER VORST, An iterative solution method for linear systems of which the coefficient matrix is a symmetric M-matrix, Mathematics of Computation, 31 (1977), pp. 148–162. [152] , Guidelines for the usage of incomplete decompositions in solving sets of linear equa- tions as they occur in practical problems, J. Comput. Phys., 44 (1981), pp. 134–155. [153] R. MELHEM, Toward efficient implementation of preconditioned conjugate gradient meth- ods on vector supercomputers, Internat. J. Supercomput. Appls., 1 (1987), pp. 77–98. [154] G. MEURANT, The block preconditioned conjugate gradient method on vector computers, BIT, 24 (1984), pp. 623–633. [155] , Multitasking the conjugate gradient method on the CRAY X-MP/48, Parallel Comput., 5 (1987), pp. 267–280. [156] N. MUNKSGAARD, Solving sparse symmetric sets of linear equations by preconditioned conjugate gradients, ACM Trans. Math. Software, 6 (1980), pp. 206–219.
- 112. 102 BIBLIOGRAPHY [157] N. NACHTIGAL, S. REDDY, AND L. TREFETHEN, How fast are nonsymmetric matrix iter- ations?, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 778–795. [158] N. NACHTIGAL, L. REICHEL, AND L. TREFETHEN, A hybrid GMRES algorithm for non- symmetric matrix iterations, Tech. Rep. 90-7, MIT, Cambridge, MA, 1990. [159] N. M. NACHTIGAL, A Look-Ahead Variant of the Lanczos Algorithm and its Application to the Quasi-Minimal Residual Methods for Non-Hermitian Linear Systems, PhD thesis, MIT, Cambridge, MA, 1991. [160] Y. NOTAY, Solving positive (semi)definite linear systems by preconditioned iterative meth- ods, in Preconditioned Conjugate Gradient Methods, O. Axelsson and L. Kolotilina, eds., vol. 1457 of Lecture Notes in Mathematics, Nijmegen, 1989, pp. 105–125. [161] , On the robustness of modified incomplete factorization methods, Internat. J. Comput. Math., 40 (1992), pp. 121–141. [162] D. O’LEARY, The block conjugate gradient algorithm and related methods, Linear Algebra Appl., 29 (1980), pp. 293–322. [163] , Ordering schemes for parallel processing of certain mesh problems, SIAM J. Sci. Statist. Comput., 5 (1984), pp. 620–632. [164] T. C. OPPE, W. D. JOUBERT, AND D. R. KINCAID, NSPCG user’s guide, version 1.0: A package for solving large sparse linear systems by various iterative methods, Tech. Rep. CNA–216, Center for Numerical Analysis, University of Texas at Austin, Austin, TX, April 1988. [165] J. M. ORTEGA, Introduction to Parallel and Vector Solution of Linear Systems, Plenum Press, New York and London, 1988. [166] C. PAIGE, B. PARLETT, AND H. VAN DER VORST, Approximate solutions and eigenvalue bounds from Krylov subspaces, Numer. Lin. Alg. Appls., to appear. [167] C. PAIGE AND M. SAUNDERS, Solution of sparse indefinite systems of linear equations, SIAM J. Numer. Anal., 12 (1975), pp. 617–629. [168] C. C. PAIGE AND M. A. SAUNDERS, LSQR: An algorithm for sparse linear equations and sparse least squares, ACM Trans. Math. Soft., 8 (1982), pp. 43–71. [169] G. PAOLINI AND G. RADICATI DI BROZOLO, Data structures to vectorize CG algorithms for general sparsity patterns, BIT, 29 (1989), pp. 703–718. [170] B. PARLETT, The symmetric eigenvalue problem, Prentice-Hall, London, 1980. [171] B. N. PARLETT, D. R. TAYLOR, AND Z. A. LIU, A look-ahead Lanczos algorithm for unsymmetric matrices, Mathematics of Computation, 44 (1985), pp. 105–124. [172] D. PEACEMAN AND J. H.H. RACHFORD, The numerical solution of parabolic and elliptic differential equations, J. Soc. Indust. Appl. Math., 3 (1955), pp. 28–41. [173] C. POMMERELL, Solution of Large Unsymmetric Systems of Linear Equations, vol. 17 of Series in Micro-electronics, volume 17, Hartung-Gorre Verlag, Konstanz, 1992. [174] , Solution of large unsymmetric systems of linear equations, PhD thesis, Swiss Federal Institute of Technology, Z¨urich, Switzerland, 1992.
- 113. BIBLIOGRAPHY 103 [175] E. POOLE AND J. ORTEGA, Multicolor ICCG methods for vector computers, Tech. Rep. RM 86-06, Department of Applied Mathematics, University of Virginia, Charlottesville, VA, 1986. [176] A. QUARTERONI, ed., Domain Decomposition Methods, Proceedings of the Sixth Interna- tional Symposium on Domain Decomposition Methods, Como, Italy,, Providence, RI, 1993, AMS. to appear. [177] G. RADICATI DI BROZOLO AND Y. ROBERT, Vector and parallel CG-like algorithms for sparse non-symmetric systems, Tech. Rep. 681-M, IMAG/TIM3, Grenoble, France, 1987. [178] J. REID, On the method of conjugate gradients for the solution of large sparse systems of linear equations, in Large Sparse Sets of Linear Equations, J. Reid, ed., Academic Press, London, 1971, pp. 231–254. [179] G. RODRIGUE AND D. WOLITZER, Preconditioning by incomplete block cyclic reduction, Mathematics of Computation, 42 (1984), pp. 549–565. [180] Y. SAAD, The Lanczos biorthogonalization algorithm and other oblique projection methods for solving large unsymmetric systems, SIAM J. Numer. Anal., 19 (1982), pp. 485–506. [181] , Practical use of some Krylov subspace methods for solving indefinite and nonsym- metric linear systems, SIAM J. Sci. Statist. Comput., 5 (1984), pp. 203–228. [182] , Practical use of polynomial preconditionings for the conjugate gradient method, SIAM J. Sci. Statist. Comput., 6 (1985), pp. 865–881. [183] , Preconditioning techniques for indefinite and nonsymmetric linear systems, J. Com- put. Appl. Math., 24 (1988), pp. 89–105. [184] , Krylov subspace methods on supercomputers, SIAM J. Sci. Statist. Comput., 10 (1989), pp. 1200–1232. [185] , SPARSKIT: A basic tool kit for sparse matrix computation, Tech. Rep. CSRD TR 1029, CSRD, University of Illinois, Urbana, IL, 1990. [186] , A flexible inner-outer preconditioned GMRES algorithm, SIAM J. Sci. Comput., 14 (1993), pp. 461–469. [187] Y. SAAD AND M. SCHULTZ, Conjugate gradient-like algorithms for solving nonsymmetric linear systems, Mathematics of Computation, 44 (1985), pp. 417–424. [188] , GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear systems, SIAM J. Sci. Statist. Comput., 7 (1986), pp. 856–869. [189] G. L. G. SLEIJPEN AND D. R. FOKKEMA, Bi-CGSTAB( ) for linear equations involving unsymmetric matrices with complex spectrum, Tech. Rep. 772, University of Utrecht, De- partment of Mathematics, Utrecht, The Netherlands, 1993. [190] B. F. SMITH, Domain decomposition algorithms for partial differential equations of linear elasticity, Tech. Rep. 517, Department of Computer Science, Courant Institute, 1990. [191] P. SONNEVELD, CGS, a fast Lanczos-type solver for nonsymmetric linear systems, SIAM J. Sci. Statist. Comput., 10 (1989), pp. 36–52. [192] R. SOUTHWELL, Relaxation Methods in Theoretical Physics, Clarendon Press, Oxford, 1946.
- 114. 104 BIBLIOGRAPHY [193] H. STONE, Iterative solution of implicit approximations of multidimensional partial differ- ential equations, SIAM J. Numer. Anal., 5 (1968), pp. 530–558. [194] P. SWARZTRAUBER, The methods of cyclic reduction, Fourier analysis and the FACR algo- rithm for the discrete solution of Poisson’s equation on a rectangle, SIAM Rev., 19 (1977), pp. 490–501. [195] C. TONG, A comparative study of preconditioned Lanczos methods for nonsymmetric linear systems, Tech. Rep. SAND91-8240, Sandia Nat. Lab., Livermore, CA, 1992. [196] A. VAN DER SLUIS, Condition numbers and equilibration of matrices, Numer. Math., 14 (1969), pp. 14–23. [197] A. VAN DER SLUIS AND H. VAN DER VORST, The rate of convergence of conjugate gradi- ents, Numer. Math., 48 (1986), pp. 543–560. [198] H. VAN DER VORST, Iterative solution methods for certain sparse linear systems with a non-symmetric matrix arising from PDE-problems, J. Comput. Phys., 44 (1981), pp. 1–19. [199] , A vectorizable variant of some ICCG methods, SIAM J. Sci. Statist. Comput., 3 (1982), pp. 350–356. [200] , Large tridiagonal and block tridiagonal linear systems on vector and parallel com- puters, Parallel Comput., 5 (1987), pp. 45–54. [201] , (M)ICCG for 2D problems on vector computers, in Supercomputing, A.Lichnewsky and C.Saguez, eds., North-Holland, 1988. Also as Report No.A-17, Data Processing Center, Kyoto University, Kyoto, Japan, December 17, 1986. [202] , High performance preconditioning, SIAM J. Sci. Statist. Comput., 10 (1989), pp. 1174–1185. [203] , ICCG and related methods for 3D problems on vector computers, Computer Physics Communications, (1989), pp. 223–235. Also as Report No.A-18, Data Processing Center, Kyoto University, Kyoto, Japan, May 30, 1987. [204] , The convergence behavior of preconditioned CG and CG-S in the presence of round- ing errors, in Preconditioned Conjugate Gradient Methods, O. Axelsson and L. Y. Kolotilina, eds., vol. 1457 of Lecture Notes in Mathematics, Berlin, New York, 1990, Springer-Verlag. [205] , Bi-CGSTAB: A fast and smoothly converging variant of Bi-CG for the solution of nonsymmetric linear systems, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 631–644. [206] H. VAN DER VORST AND J. MELISSEN, A Petrov-Galerkin type method for solving Ax = b where A is symmetric complex, IEEE Trans. Magnetics, 26 (1990), pp. 706–708. [207] H. VAN DER VORST AND C. VUIK, GMRESR: A family of nested GMRES methods, Tech. Rep. 91-80, Delft University of Technology, Faculty of Tech. Math., Delft, The Netherlands, 1991. [208] J. VAN ROSENDALE, Minimizing inner product data dependencies in conjugate gradient iteration, Tech. Rep. 172178, ICASE, NASA Langley Research Center, 1983. [209] R. VARGA, Matrix Iterative Analysis, Prentice-Hall Inc., Englewood Cliffs, NJ, 1962. [210] P. VASSILEVSKI, Preconditioning nonsymmetric and indefinite finite element matrices, J. Numer. Alg. Appl., 1 (1992), pp. 59–76.
- 115. BIBLIOGRAPHY 105 [211] V. VOEVODIN, The problem of non-self-adjoint generalization of the conjugate gradient method is closed, U.S.S.R. Comput. Maths. and Math. Phys., 23 (1983), pp. 143–144. [212] H. F. WALKER, Implementation of the GMRES method using Householder transformations, SIAM J. Sci. Statist. Comput., 9 (1988), pp. 152–163. [213] P. WESSELING, An Introduction to Multigrid Methods, Wiley, Chichester, 1991. [214] O. WIDLUND, A Lanczos method for a class of non-symmetric systems of linear equations, SIAM J. Numer. Anal., 15 (1978), pp. 801–812. [215] D. YOUNG, Iterative solution of large linear systems, Academic Press, New York, 1971. [216] H. YSERENTANT, On the multilevel splitting of finite element spaces, Numer. Math., 49 (1986), pp. 379–412.
- 116. 106 BIBLIOGRAPHY
- 117. Appendix D Flowchart of iterative methods y n y n y n y n y n Is the transpose available? Is storage at a premium? Try CG Is the matrix symmetric? Is the matrix definite? Are the outer eigenvalues known? Try Chebyshev or CG Try GMRES with long restart Try MinRES or CG Try QMR Try CGS or Bi-CGSTAB 107