Iterative Methods for Linear

and Nonlinear Equations
C. T. Kelley
North Carolina State University
Society for Industrial and Applied Mathematics
Philadelphia 1995
Untitled-1 9/20/2004, 2:59 PM 3
To Polly H. Thomas, 1906-1994, devoted mother and grandmother
1
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
2
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Contents
Preface xi
How to Get the Software xiii
CHAPTER 1. Basic Concepts and Stationary Iterative Methods 3
1.1 Review and notation . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The Banach Lemma and approximate inverses . . . . . . . . . . 5
1.3 The spectral radius . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Matrix splittings and classical stationary iterative methods . . 7
1.5 Exercises on stationary iterative methods . . . . . . . . . . . . 10
CHAPTER 2. Conjugate Gradient Iteration 11
2.1 Krylov methods and the minimization property . . . . . . . . . 11
2.2 Consequences of the minimization property . . . . . . . . . . . 13
2.3 Termination of the iteration . . . . . . . . . . . . . . . . . . . . 15
2.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6 CGNR and CGNE . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 Examples for preconditioned conjugate iteration . . . . . . . . 26
2.8 Exercises on conjugate gradient . . . . . . . . . . . . . . . . . . 30
CHAPTER 3. GMRES Iteration 33
3.1 The minimization property and its consequences . . . . . . . . 33
3.2 Termination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 GMRES implementation: Basic ideas . . . . . . . . . . . . . . . 37
3.5 Implementation: Givens rotations . . . . . . . . . . . . . . . . . 43
3.6 Other methods for nonsymmetric systems . . . . . . . . . . . . 46
3.6.1 Bi-CG. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6.2 CGS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6.3 Bi-CGSTAB. . . . . . . . . . . . . . . . . . . . . . . . . 50
vii
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
viii CONTENTS
3.6.4 TFQMR. . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.7 Examples for GMRES iteration . . . . . . . . . . . . . . . . . . 54
3.8 Examples for CGNR, Bi-CGSTAB, and TFQMR iteration . . . 55
3.9 Exercises on GMRES . . . . . . . . . . . . . . . . . . . . . . . . 60
CHAPTER 4. Basic Concepts and Fixed-Point Iteration 65
4.1 Types of convergence . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Fixed-point iteration . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 The standard assumptions . . . . . . . . . . . . . . . . . . . . . 68
CHAPTER 5. Newton’s Method 71
5.1 Local convergence of Newton’s method . . . . . . . . . . . . . . 71
5.2 Termination of the iteration . . . . . . . . . . . . . . . . . . . . 72
5.3 Implementation of Newton’s method . . . . . . . . . . . . . . . 73
5.4 Errors in the function and derivative . . . . . . . . . . . . . . . 75
5.4.1 The chord method. . . . . . . . . . . . . . . . . . . . . . 76
5.4.2 Approximate inversion of F

. . . . . . . . . . . . . . . . 77
5.4.3 The Shamanskii method. . . . . . . . . . . . . . . . . . 78
5.4.4 Difference approximation to F

. . . . . . . . . . . . . . . 79
5.4.5 The secant method. . . . . . . . . . . . . . . . . . . . . 82
5.5 The Kantorovich Theorem . . . . . . . . . . . . . . . . . . . . . 83
5.6 Examples for Newton’s method . . . . . . . . . . . . . . . . . . 86
5.7 Exercises on Newton’s method . . . . . . . . . . . . . . . . . . 91
CHAPTER 6. Inexact Newton Methods 95
6.1 The basic estimates . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.1.1 Direct analysis. . . . . . . . . . . . . . . . . . . . . . . . 95
6.1.2 Weighted norm analysis. . . . . . . . . . . . . . . . . . . 97
6.1.3 Errors in the function. . . . . . . . . . . . . . . . . . . . 100
6.2 Newton-iterative methods . . . . . . . . . . . . . . . . . . . . . 100
6.2.1 Newton GMRES. . . . . . . . . . . . . . . . . . . . . . . 101
6.2.2 Other Newton-iterative methods. . . . . . . . . . . . . . 104
6.3 Newton-GMRES implementation . . . . . . . . . . . . . . . . . 104
6.4 Examples for Newton-GMRES . . . . . . . . . . . . . . . . . . 106
6.4.1 Chandrasekhar H-equation. . . . . . . . . . . . . . . . . 107
6.4.2 Convection-diffusion equation. . . . . . . . . . . . . . . 108
6.5 Exercises on inexact Newton methods . . . . . . . . . . . . . . 110
CHAPTER 7. Broyden’s method 113
7.1 The Dennis–Mor´e condition . . . . . . . . . . . . . . . . . . . . 114
7.2 Convergence analysis . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2.1 Linear problems. . . . . . . . . . . . . . . . . . . . . . . 118
7.2.2 Nonlinear problems. . . . . . . . . . . . . . . . . . . . . 120
7.3 Implementation of Broyden’s method . . . . . . . . . . . . . . . 123
7.4 Examples for Broyden’s method . . . . . . . . . . . . . . . . . . 127
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONTENTS ix
7.4.1 Linear problems. . . . . . . . . . . . . . . . . . . . . . . 127
7.4.2 Nonlinear problems. . . . . . . . . . . . . . . . . . . . . 128
7.5 Exercises on Broyden’s method . . . . . . . . . . . . . . . . . . 132
CHAPTER 8. Global Convergence 135
8.1 Single equations . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.2 Analysis of the Armijo rule . . . . . . . . . . . . . . . . . . . . 138
8.3 Implementation of the Armijo rule . . . . . . . . . . . . . . . . 141
8.3.1 Polynomial line searches. . . . . . . . . . . . . . . . . . 142
8.3.2 Broyden’s method. . . . . . . . . . . . . . . . . . . . . . 144
8.4 Examples for Newton–Armijo . . . . . . . . . . . . . . . . . . . 146
8.4.1 Inverse tangent function. . . . . . . . . . . . . . . . . . 146
8.4.2 Convection-diffusion equation. . . . . . . . . . . . . . . 146
8.4.3 Broyden–Armijo. . . . . . . . . . . . . . . . . . . . . . . 148
8.5 Exercises on global convergence . . . . . . . . . . . . . . . . . . 151
Bibliography 153
Index 163
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
x CONTENTS
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Preface
This book on iterative methods for linear and nonlinear equations can be used
as a tutorial and a reference by anyone who needs to solve nonlinear systems
of equations or large linear systems. It may also be used as a textbook for
introductory courses in nonlinear equations or iterative methods or as source
material for an introductory course in numerical analysis at the graduate level.
We assume that the reader is familiar with elementary numerical analysis,
linear algebra, and the central ideas of direct methods for the numerical
solution of dense linear systems as described in standard texts such as [7],
[105], or [184].
Our approach is to focus on a small number of methods and treat them
in depth. Though this book is written in a finite-dimensional setting, we
have selected for coverage mostly algorithms and methods of analysis which
extend directly to the infinite-dimensional case and whose convergence can be
thoroughly analyzed. For example, the matrix-free formulation and analysis for
GMRES and conjugate gradient is almost unchanged in an infinite-dimensional
setting. The analysis of Broyden’s method presented in Chapter 7 and
the implementations presented in Chapters 7 and 8 are different from the
classical ones and also extend directly to an infinite-dimensional setting. The
computational examples and exercises focus on discretizations of infinite-
dimensional problems such as integral and differential equations.
We present a limited number of computational examples. These examples
are intended to provide results that can be used to validate the reader’s own
implementations and to give a sense of how the algorithms perform. The
examples are not designed to give a complete picture of performance or to be
a suite of test problems.
The computational examples in this book were done with MATLAB

(version 4.0a on various SUN SPARCstations and version 4.1 on an Apple
Macintosh Powerbook 180) and the MATLAB environment is an excellent one
for getting experience with the algorithms, for doing the exercises, and for
small-to-medium scale production work.
1
MATLAB codes for many of the
algorithms are available by anonymous ftp. A good introduction to the latest
1
MATLAB is a registered trademark of The MathWorks, Inc.
xi
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
xii PREFACE
version (version 4.2) of MATLAB is the MATLAB Primer [178]; [43] is also
a useful resource. If the reader has no access to MATLAB or will be solving
very large problems, the general algorithmic descriptions or even the MATLAB
codes can easily be translated to another language.
Parts of this book are based upon work supported by the National
Science Foundation and the Air Force Office of Scientific Research over
several years, most recently under National Science Foundation Grant Nos.
DMS-9024622 and DMS-9321938. Any opinions, findings, and conclusions or
recommendations expressed in this material are those of the author and do not
necessarily reflect the views of the National Science Foundation or of the Air
Force Office of Scientific Research.
Many of my students and colleagues discussed various aspects of this
project with me and provided important corrections, ideas, suggestions, and
pointers to the literature. I am especially indebted to Jim Banoczi, Jeff Butera,
Steve Campbell, Tony Choi, Moody Chu, Howard Elman, Jim Epperson,
Andreas Griewank, Laura Helfrich, Ilse Ipsen, Lea Jenkins, Vickie Kearn,
Belinda King, Debbie Lockhart, Carl Meyer, Casey Miller, Ekkehard Sachs,
Jeff Scroggs, Joseph Skudlarek, Mike Tocci, Gordon Wade, Homer Walker,
Steve Wright, Zhaqing Xue, Yue Zhang, and an anonymous reviewer for their
contributions and encouragement.
Most importantly, I thank Chung-Wei Ng and my parents for over one
hundred and ten years of patience and support.
C. T. Kelley
Raleigh, North Carolina
January, 1998
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
How to get the software
A collection of MATLAB codes has been written to accompany this book. The
MATLAB codes can be obtained by anonymous ftp from the MathWorks server
ftp.mathworks.com in the directory pub/books/kelley, from the MathWorks
World Wide Web site,
http://www.mathworks.com
or from SIAM’s World Wide Web site
http://www.siam.org/books/kelley/kelley.html
One can obtain MATLAB from
The MathWorks, Inc.
24 Prime Park Way
Natick, MA 01760,
Phone: (508) 653-1415
Fax: (508) 653-2997
E-mail: info@mathworks.com
WWW: http://www.mathworks.com
xiii
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 1
Basic Concepts and Stationary Iterative Methods
1.1. Review and notation
We begin by setting notation and reviewing some ideas from numerical linear
algebra that we expect the reader to be familiar with. An excellent reference
for the basic ideas of numerical linear algebra and direct methods for linear
equations is [184].
We will write linear equations as
Ax = b, (1.1)
where A is a nonsingular N N matrix, b ∈ R
N
is given, and
x

= A
−1
b ∈ R
N
is to be found.
Throughout this chapter x will denote a potential solution and ¦x
k
¦
k≥0
the
sequence of iterates. We will denote the ith component of a vector x by (x)
i
(note the parentheses) and the ith component of x
k
by (x
k
)
i
. We will rarely
need to refer to individual components of vectors.
In this chapter | | will denote a norm on R
N
as well as the induced matrix
norm.
Definition 1.1.1. Let | | be a norm on R
N
. The induced matrix norm
of an N N matrix A is defined by
|A| = max
x=1
|Ax|.
Induced norms have the important property that
|Ax| ≤ |A||x|.
Recall that the condition number of A relative to the norm | | is
κ(A) = |A||A
−1
|,
where κ(A) is understood to be infinite if A is singular. If | | is the l
p
norm
|x|
p
=
_
_
N

j=1
[(x)
i
[
p
_
_
1/p
3
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
4 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
we will write the condition number as κ
p
.
Most iterative methods terminate when the residual
r = b −Ax
is sufficiently small. One termination criterion is
|r
k
|
|r
0
|
< τ, (1.2)
which can be related to the error
e = x −x

in terms of the condition number.
Lemma 1.1.1. Let b, x, x
0
∈ R
N
. Let A be nonsingular and let x

= A
−1
b.
|e|
|e
0
|
≤ κ(A)
|r|
|r
0
|
. (1.3)
Proof. Since
r = b −Ax = −Ae
we have
|e| = |A
−1
Ae| ≤ |A
−1
||Ae| = |A
−1
||r|
and
|r
0
| = |Ae
0
| ≤ |A||e
0
|.
Hence
|e|
|e
0
|

|A
−1
||r|
|A|
−1
|r
0
|
= κ(A)
|r|
|r
0
|
,
as asserted.
The termination criterion (1.2) depends on the initial iterate and may result
in unnecessary work when the initial iterate is good and a poor result when the
initial iterate is far from the solution. For this reason we prefer to terminate
the iteration when
|r
k
|
|b|
< τ. (1.4)
The two conditions (1.2) and (1.4) are the same when x
0
= 0, which is a
common choice, particularly when the linear iteration is being used as part of
a nonlinear solver.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BASIC CONCEPTS 5
1.2. The Banach Lemma and approximate inverses
The most straightforward approach to an iterative solution of a linear system
is to rewrite (1.1) as a linear fixed-point iteration. One way to do this is to
write Ax = b as
x = (I −A)x +b, (1.5)
and to define the Richardson iteration
x
k+1
= (I −A)x
k
+b. (1.6)
We will discuss more general methods in which ¦x
k
¦ is given by
x
k+1
= Mx
k
+c. (1.7)
In (1.7) M is an NN matrix called the iteration matrix. Iterative methods of
this form are called stationary iterative methods because the transition from x
k
to x
k+1
does not depend on the history of the iteration. The Krylov methods
discussed in Chapters 2 and 3 are not stationary iterative methods.
All our results are based on the following lemma.
Lemma 1.2.1. If M is an N N matrix with |M| < 1 then I − M is
nonsingular and
|(I −M)
−1
| ≤
1
1 −|M|
. (1.8)
Proof. We will show that I − M is nonsingular and that (1.8) holds by
showing that the series

l=0
M
l
= (I −M)
−1
.
The partial sums
S
k
=
k

l=0
M
l
form a Cauchy sequence in R
N×N
. To see this note that for all m > k
|S
k
−S
m
| ≤
m

l=k+1
|M
l
|.
Now, |M
l
| ≤ |M|
l
because | | is a matrix norm that is induced by a vector
norm. Hence
|S
k
−S
m
| ≤
m

l=k+1
|M|
l
= |M|
k+1
_
1 −|M|
m−k
1 −|M|
_
→ 0
as m, k → ∞. Hence the sequence S
k
converges, say to S. Since MS
k
+ I =
S
k+1
, we must have MS +I = S and hence (I −M)S = I. This proves that
I −M is nonsingular and that S = (I −M)
−1
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
6 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Noting that
|(I −M)
−1
| ≤

l=0
|M|
l
= (1 −|M|)
−1
.
proves (1.8) and completes the proof.
The following corollary is a direct consequence of Lemma 1.2.1.
Corollary 1.2.1. If |M| < 1 then the iteration (1.7) converges to
x = (I −M)
−1
c for all initial iterates x
0
.
A consequence of Corollary 1.2.1 is that Richardson iteration (1.6) will
converge if |I − A| < 1. It is sometimes possible to precondition a linear
equation by multiplying both sides of (1.1) by a matrix B
BAx = Bb
so that convergence of iterative methods is improved. In the context of
Richardson iteration, the matrices B that allow us to apply the Banach lemma
and its corollary are called approximate inverses.
Definition 1.2.1. B is an approximate inverse of A if |I −BA| < 1.
The following theorem is often referred to as the Banach Lemma.
Theorem 1.2.1. If A and B are NN matrices and B is an approximate
inverse of A, then A and B are both nonsingular and
|A
−1
| ≤
|B|
1 −|I −BA|
, |B
−1
| ≤
|A|
1 −|I −BA|
, (1.9)
and
|A
−1
−B| ≤
|B||I −BA|
1 −|I −BA|
, |A−B
−1
| ≤
|A||I −BA|
1 −|I −BA|
. (1.10)
Proof. Let M = I −BA. By Lemma 1.2.1 I −M = I −(I −BA) = BA is
nonsingular. Hence both A and B are nonsingular. By (1.8)
|A
−1
B
−1
| = |(I −M)
−1
| ≤
1
1 −|M|
=
1
1 −|I −BA|
. (1.11)
Since A
−1
= (I −M)
−1
B, inequality (1.11) implies the first part of (1.9). The
second part follows in a similar way from B
−1
= A(I −M)
−1
.
To complete the proof note that
A
−1
−B = (I −BA)A
−1
, A−B
−1
= B
−1
(I −BA),
and use (1.9).
Richardson iteration, preconditioned with approximate inversion, has the
form
x
k+1
= (I −BA)x
k
+Bb. (1.12)
If the norm of I − BA is small, then not only will the iteration converge
rapidly, but, as Lemma 1.1.1 indicates, termination decisions based on the
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BASIC CONCEPTS 7
preconditioned residual Bb − BAx will better reflect the actual error. This
method is a very effective technique for solving differential equations, integral
equations, and related problems [15], [6], [100], [117], [111]. Multigrid methods
[19], [99], [126], can also be interpreted in this light. We mention one other
approach, polynomial preconditioning, which tries to approximate A
−1
by a
polynomial in A [123], [179], [169].
1.3. The spectral radius
The analysis in ¸ 1.2 related convergence of the iteration (1.7) to the norm of
the matrix M. However the norm of M could be small in some norms and
quite large in others. Hence the performance of the iteration is not completely
described by |M|. The concept of spectral radius allows us to make a complete
description.
We let σ(A) denote the set of eigenvalues of A.
Definition 1.3.1. The spectral radius of an N N matrix A is
ρ(A) = max
λ∈σ(A)
[λ[ = lim
n→∞
|A
n
|
1/n
. (1.13)
The term on the right-hand side of the second equality in (1.13) is the limit
used by the radical test for convergence of the series

A
n
.
The spectral radius of M is independent of any particular matrix norm of
M. It is clear, in fact, that
ρ(A) ≤ |A| (1.14)
for any induced matrix norm. The inequality (1.14) has a partial converse that
allows us to completely describe the performance of iteration (1.7) in terms of
spectral radius. We state that converse as a theorem and refer to [105] for a
proof.
Theorem 1.3.1. Let A be an N N matrix. Then for any > 0 there is
a norm | | on R
N
such that
ρ(A) > |A| −.
A consequence of Theorem 1.3.1, Lemma 1.2.1, and Exercise 1.5.1 is a
characterization of convergent stationary iterative methods. The proof is left
as an exercise.
Theorem 1.3.2. Let M be an NN matrix. The iteration (1.7) converges
for all c ∈ R
N
if and only if ρ(M) < 1.
1.4. Matrix splittings and classical stationary iterative methods
There are ways to convert Ax = b to a linear fixed-point iteration that are
different from (1.5). Methods such as Jacobi, Gauss–Seidel, and sucessive
overrelaxation (SOR) iteration are based on splittings of A of the form
A = A
1
+A
2
,
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
8 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
where A
1
is a nonsingular matrix constructed so that equations with A
1
as
coefficient matrix are easy to solve. Then Ax = b is converted to the fixed-
point problem
x = A
−1
1
(b −A
2
x).
The analysis of the method is based on an estimation of the spectral radius of
the iteration matrix M = −A
−1
1
A
2
.
For a detailed description of the classical stationary iterative methods the
reader may consult [89], [105], [144], [193], or [200]. These methods are usually
less efficient than the Krylov methods discussed in Chapters 2 and 3 or the more
modern stationary methods based on multigrid ideas. However the classical
methods have a role as preconditioners. The limited description in this section
is intended as a review that will set some notation to be used later.
As a first example we consider the Jacobi iteration that uses the splitting
A
1
= D, A
2
= L +U,
where D is the diagonal of A and L and U are the (strict) lower and upper
triangular parts. This leads to the iteration matrix
M
JAC
= −D
−1
(L +U).
Letting (x
k
)
i
denote the ith component of the kth iterate we can express Jacobi
iteration concretely as
(x
k+1
)
i
= a
−1
ii
_
_
b
i

j=i
a
ij
(x
k
)
j
_
_
. (1.15)
Note that A
1
is diagonal and hence trivial to invert.
We present only one convergence result for the classical stationary iterative
methods.
Theorem 1.4.1. Let A be an N N matrix and assume that for all
1 ≤ i ≤ N
0 <

j=i
[a
ij
[ < [a
ii
[. (1.16)
Then A is nonsingular and the Jacobi iteration (1.15) converges to x

= A
−1
b
for all b.
Proof. Note that the ith row sum of M = M
JAC
satisfies
N

j=1
[m
ij
[ =

j=i
[a
ij
[
[a
ii
[
< 1.
Hence |M
JAC
|

< 1 and the iteration converges to the unique solution of
x = Mx + D
−1
b. Also I − M = D
−1
A is nonsingular and therefore A is
nonsingular.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BASIC CONCEPTS 9
Gauss–Seidel iteration overwrites the approximate solution with the new
value as soon as it is computed. This results in the iteration
(x
k+1
)
i
= a
−1
ii
_
_
b
i

j<i
a
ij
(x
k+1
)
j

j>i
a
ij
(x
k
)
j
_
_
,
the splitting
A
1
= D +L, A
2
= U,
and iteration matrix
M
GS
= −(D +L)
−1
U.
Note that A
1
is lower triangular, and hence A
−1
1
y is easy to compute for vectors
y. Note also that, unlike Jacobi iteration, the iteration depends on the ordering
of the unknowns. Backward Gauss–Seidel begins the update of x with the Nth
coordinate rather than the first, resulting in the splitting
A
1
= D +U, A
2
= L,
and iteration matrix
M
BGS
= −(D +U)
−1
L.
A symmetric Gauss–Seidel iteration is a forward Gauss–Seidel iteration
followed by a backward Gauss–Seidel iteration. This leads to the iteration
matrix
M
SGS
= M
BGS
M
GS
= (D +U)
−1
L(D +L)
−1
U.
If A is symmetric then U = L
T
. In that event
M
SGS
= (D +U)
−1
L(D +L)
−1
U = (D +L
T
)
−1
L(D +L)
−1
L
T
.
From the point of view of preconditioning, one wants to write the stationary
method as a preconditioned Richardson iteration. That means that one wants
to find B such that M = I − BA and then use B as an approximate inverse.
For the Jacobi iteration,
B
JAC
= D
−1
. (1.17)
For symmetric Gauss–Seidel
B
SGS
= (D +L
T
)
−1
D(D +L)
−1
. (1.18)
The successive overrelaxation iteration modifies Gauss–Seidel by adding a
relaxation parameter ω to construct an iteration with iteration matrix
M
SOR
= (D +ωL)
−1
((1 −ω)D −ωU).
The performance can be dramatically improved with a good choice of ω but
still is not competitive with Krylov methods. A further disadvantage is that
the choice of ω is often difficult to make. References [200], [89], [193], [8], and
the papers cited therein provide additional reading on this topic.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
10 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
1.5. Exercises on stationary iterative methods
1.5.1. Show that if ρ(M) ≥ 1 then there are x
0
and c such that the iteration
(1.7) fails to converge.
1.5.2. Prove Theorem 1.3.2.
1.5.3. Verify equality (1.18).
1.5.4. Show that if A is symmetric and positive definite (that is A
T
= A and
x
T
Ax > 0 for all x ,= 0) that B
SGS
is also symmetric and positive
definite.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 2
Conjugate Gradient Iteration
2.1. Krylov methods and the minimization property
In the following two chapters we describe some of the Krylov space methods
for linear equations. Unlike the stationary iterative methods, Krylov methods
do not have an iteration matrix. The two such methods that we’ll discuss in
depth, conjugate gradient and GMRES, minimize, at the kth iteration, some
measure of error over the affine space
x
0
+/
k
,
where x
0
is the initial iterate and the kth Krylov subspace /
k
is
/
k
= span(r
0
, Ar
0
, . . . , A
k−1
r
0
)
for k ≥ 1.
The residual is
r = b −Ax.
So ¦r
k
¦
k≥0
will denote the sequence of residuals
r
k
= b −Ax
k
.
As in Chapter 1, we assume that A is a nonsingular N N matrix and let
x

= A
−1
b.
There are other Krylov methods that are not as well understood as CG or
GMRES. Brief descriptions of several of these methods and their properties
are in ¸ 3.6, [12], and [78].
The conjugate gradient (CG) iteration was invented in the 1950s [103] as a
direct method. It has come into wide use over the last 15 years as an iterative
method and has generally superseded the Jacobi–Gauss–Seidel–SOR family of
methods.
CG is intended to solve symmetric positive definite (spd) systems. Recall
that A is symmetric if A = A
T
and positive definite if
x
T
Ax > 0 for all x ,= 0.
11
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
12 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
In this section we assume that A is spd. Since A is spd we may define a norm
(you should check that this is a norm) by
|x|
A
=

x
T
Ax. (2.1)
| |
A
is called the A-norm. The development in these notes is different from
the classical work and more like the analysis for GMRES and CGNR in [134].
In this section, and in the section on GMRES that follows, we begin with a
description of what the algorithm does and the consequences of the minimiza-
tion property of the iterates. After that we describe termination criterion,
performance, preconditioning, and at the very end, the implementation.
The kth iterate x
k
of CG minimizes
φ(x) =
1
2
x
T
Ax −x
T
b (2.2)
over x
0
+/
k
.
Note that if φ(˜ x) is the minimal value (in R
N
) then
∇φ(˜ x) = A˜ x −b = 0
and hence ˜ x = x

.
Minimizing φ over any subset of R
N
is the same as minimizing |x −x

|
A
over that subset. We state this as a lemma.
Lemma 2.1.1. Let o ⊂ R
N
. If x
k
minimizes φ over o then x
k
also
minimizes |x

−x|
A
= |r|
A
−1 over o.
Proof. Note that
|x −x

|
2
A
= (x −x

)
T
A(x −x

) = x
T
Ax −x
T
Ax

−(x

)
T
Ax + (x

)
T
Ax

.
Since A is symmetric and Ax

= b
−x
T
Ax

−(x

)
T
Ax = −2x
T
Ax

= −2x
T
b.
Therefore
|x −x

|
2
A
= 2φ(x) + (x

)
T
Ax

.
Since (x

)
T
Ax

is independent of x, minimizing φ is equivalent to minimizing
|x −x

|
2
A
and hence to minimizing |x −x

|
A
.
If e = x −x

then
|e|
2
A
= e
T
Ae = (A(x −x

))
T
A
−1
(A(x −x

)) = |b −Ax|
2
A
−1
and hence the A-norm of the error is also the A
−1
-norm of the residual.
We will use this lemma in the particular case that o = x
0
+/
k
for some k.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 13
2.2. Consequences of the minimization property
Lemma 2.1.1 implies that since x
k
minimizes φ over x
0
+/
k
|x

−x
k
|
A
≤ |x

−w|
A
(2.3)
for all w ∈ x
0
+/
k
. Since any w ∈ x
0
+/
k
can be written as
w =
k−1

j=0
γ
j
A
j
r
0
+x
0
for some coefficients ¦γ
j
¦, we can express x

−w as
x

−w = x

−x
0

k−1

j=0
γ
j
A
j
r
0
.
Since Ax

= b we have
r
0
= b −Ax
0
= A(x

−x
0
)
and therefore
x

−w = x

−x
0

k−1

j=0
γ
j
A
j+1
(x

−x
0
) = p(A)(x

−x
0
),
where the polynomial
p(z) = 1 −
k−1

j=0
γ
j
z
j+1
has degree k and satisfies p(0) = 1. Hence
|x

−x
k
|
A
= min
p∈P
k
,p(0)=1
|p(A)(x

−x
0
)|
A
. (2.4)
In (2.4) P
k
denotes the set of polynomials of degree k.
The spectral theorem for spd matrices asserts that
A = UΛU
T
,
where U is an orthogonal matrix whose columns are the eigenvectors of A and
Λ is a diagonal matrix with the positive eigenvalues of A on the diagonal. Since
UU
T
= U
T
U = I by orthogonality of U, we have
A
j
= UΛ
j
U
T
.
Hence
p(A) = Up(Λ)U
T
.
Define A
1/2
= UΛ
1/2
U
T
and note that
|x|
2
A
= x
T
Ax = |A
1/2
x|
2
2
. (2.5)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
14 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Hence, for any x ∈ R
N
and
|p(A)x|
A
= |A
1/2
p(A)x|
2
≤ |p(A)|
2
|A
1/2
x|
2
≤ |p(A)|
2
|x|
A
.
This, together with (2.4) implies that
|x
k
−x

|
A
≤ |x
0
−x

|
A
min
p∈P
k
,p(0)=1
max
z∈σ(A)
[p(z)[. (2.6)
Here σ(A) is the set of all eigenvalues of A.
The following corollary is an important consequence of (2.6).
Corollary 2.2.1. Let A be spd and let ¦x
k
¦ be the CG iterates. Let k be
given and let ¦¯ p
k
¦ be any kth degree polynomial such that ¯ p
k
(0) = 1. Then
|x
k
−x

|
A
|x
0
−x

|
A
≤ max
z∈σ(A)
[ ¯ p
k
(z)[. (2.7)
We will refer to the polynomial ¯ p
k
as a residual polynomial [185].
Definition 2.2.1. The set of kth degree residual polynomials is
T
k
= ¦p [ p is a polynomial of degree k and p(0) = 1.¦ (2.8)
In specific contexts we try to construct sequences of residual polynomials,
based on information on σ(A), that make either the middle or the right term
in (2.7) easy to evaluate. This leads to an upper estimate for the number of
CG iterations required to reduce the A-norm of the error to a given tolerance.
One simple application of (2.7) is to show how the CG algorithm can be
viewed as a direct method.
Theorem 2.2.1. Let A be spd. Then the CG algorithm will find the
solution within N iterations.
Proof. Let ¦λ
i
¦
N
i=1
be the eigenvalues of A. As a test polynomial, let
¯ p(z) =
N

i=1

i
−z)/λ
i
.
¯ p ∈ T
N
because ¯ p has degree N and ¯ p(0) = 1. Hence, by (2.7) and the fact
that ¯ p vanishes on σ(A),
|x
N
−x

|
A
≤ |x
0
−x

|
A
max
z∈σ(A)
[ ¯ p(z)[ = 0.
Note that our test polynomial had the eigenvalues of A as its roots. In
that way we showed (in the absence of all roundoff error!) that CG terminated
in finitely many iterations with the exact solution. This is not as good as it
sounds, since in most applications the number of unknowns N is very large,
and one cannot afford to perform N iterations. It is best to regard CG as an
iterative method. When doing that we seek to terminate the iteration when
some specified error tolerance is reached.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 15
In the two examples that follow we look at some other easy consequences
of (2.7).
Theorem 2.2.2. Let A be spd with eigenvectors ¦u
i
¦
N
i=1
. Let b be a linear
combination of k of the eigenvectors of A
b =
k

l=1
γ
l
u
i
l
.
Then the CG iteration for Ax = b with x
0
= 0 will terminate in at most k
iterations.
Proof. Let ¦λ
i
l
¦ be the eigenvalues of A associated with the eigenvectors
¦u
i
l
¦
k
l=1
. By the spectral theorem
x

=
k

l=1

l

i
l
)u
i
l
.
We use the residual polynomial,
¯ p(z) =
k

l=1

i
l
−z)/λ
i
l
.
One can easily verify that ¯ p ∈ T
k
. Moreover, ¯ p(λ
i
l
) = 0 for 1 ≤ l ≤ k and
hence
¯ p(A)x

=
k

l=1
¯ p(λ
i
l

l

i
l
u
i
l
= 0.
So, we have by (2.4) and the fact that x
0
= 0 that
|x
k
−x

|
A
≤ |¯ p(A)x

|
A
= 0.
This completes the proof.
If the spectrum of A has fewer than N points, we can use a similar technique
to prove the following theorem.
Theorem 2.2.3. Let A be spd. Assume that there are exactly k ≤ N
distinct eigenvalues of A. Then the CG iteration terminates in at most k
iterations.
2.3. Termination of the iteration
In practice we do not run the CG iteration until an exact solution is found, but
rather terminate once some criterion has been satisfied. One typical criterion is
small (say ≤ η) relative residuals. This means that we terminate the iteration
after
|b −Ax
k
|
2
≤ η|b|
2
. (2.9)
The error estimates that come from the minimization property, however, are
based on (2.7) and therefore estimate the reduction in the relative A-norm of
the error.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
16 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Our next task is to relate the relative residual in the Euclidean norm to
the relative error in the A-norm. We will do this in the next two lemmas and
then illustrate the point with an example.
Lemma 2.3.1. Let A be spd with eigenvalues λ
1
≥ λ
2
≥ . . . λ
N
. Then for
all z ∈ R
N
,
|A
1/2
z|
2
= |z|
A
(2.10)
and
λ
1/2
N
|z|
A
≤ |Az|
2
≤ λ
1/2
1
|z|
A
. (2.11)
Proof. Clearly
|z|
2
A
= z
T
Az = (A
1/2
z)
T
(A
1/2
z) = |A
1/2
z|
2
2
which proves (2.10).
Let u
i
be a unit eigenvector corresponding to λ
i
. We may write A = UΛU
T
as
Az =
N

i=1
λ
i
(u
T
i
z)u
i
.
Hence
λ
N
|A
1/2
z|
2
2
= λ
N

N
i=1
λ
i
(u
T
i
z)
2
≤ |Az|
2
2
=

N
i=1
λ
2
i
(u
T
i
z)
2
≤ λ
1

N
i=1
λ
i
(u
T
i
z)
2
= λ
1
|A
1/2
z|
2
2
.
Taking square roots and using (2.10) complete the proof.
Lemma 2.3.2.
|b|
2
|r
0
|
2
|b −Ax
k
|
2
|b|
2
=
|b −Ax
k
|
2
|b −Ax
0
|
2

_
κ
2
(A)
|x
k
−x

|
A
|x

−x
0
|
A
(2.12)
and
|b −Ax
k
|
2
|b|
2

_
κ
2
(A)|r
0
|
2
|b|
2
|x
k
−x

|
A
|x

−x
0
|
A
. (2.13)
Proof. The equality on the left of (2.12) is clear and (2.13) follows directly
from (2.12). To obtain the inequality on the right of (2.12), first recall that if
A = UΛU
T
is the spectral decomposition of A and we order the eigenvalues
such that λ
1
≥ λ
2
≥ . . . λ
N
> 0, then |A|
2
= λ
1
and |A
−1
|
2
= 1/λ
N
. So
κ
2
(A) = λ
1

N
.
Therefore, using (2.10) and (2.11) twice,
|b −Ax
k
|
2
|b −Ax
0
|
2
=
|A(x

−x
k
)|
2
|A(x

−x
0
)|
2

¸
λ
1
λ
N
|x

−x
k
|
A
|x

−x
0
|
A
as asserted.
So, to predict the performance of the CG iteration based on termination on
small relative residuals, we must not only use (2.7) to predict when the relative
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 17
A-norm error is small, but also use Lemma 2.3.2 to relate small A-norm errors
to small relative residuals.
We consider a very simple example. Assume that x
0
= 0 and that the
eigenvalues of A are contained in the interval (9, 11). If we let ¯ p
k
(z) =
(10 −z)
k
/10
k
, then ¯ p
k
∈ T
k
. This means that we may apply (2.7) to get
|x
k
−x

|
A
≤ |x

|
A
max
9≤z≤11
[ ¯ p
k
(z)[.
It is easy to see that
max
9≤z≤11
[ ¯ p
k
(z)[ = 10
−k
.
Hence, after k iterations
|x
k
−x

|
A
≤ |x

|
A
10
−k
. (2.14)
So, the size of the A-norm of the error will be reduced by a factor of 10
−3
when
10
−k
≤ 10
−3
,
that is, when
k ≥ 3.
To use Lemma 2.3.2 we simply note that κ
2
(A) ≤ 11/9. Hence, after k
iterations we have
|Ax
k
−b|
2
|b|
2


11 10
−k
/3.
So, the size of the relative residual will be reduced by a factor of 10
−3
when
10
−k
≤ 3 10
−3
/

11,
that is, when
k ≥ 4.
One can obtain a more precise estimate by using a polynomial other than
p
k
in the upper estimate for the right-hand side of (2.7). Note that it is always
the case that the spectrum of a spd matrix is contained in the interval [λ
N
, λ
1
]
and that κ
2
(A) = λ
1

N
. A result from [48] (see also [45]) that is, in one
sense, the sharpest possible, is
|x
k
−x

|
A
≤ 2|x
0
−x

|
A
_
_
κ
2
(A) −1
_
κ
2
(A) + 1
_
k
. (2.15)
In the case of the above example, we can estimate κ
2
(A) by κ
2
(A) ≤ 11/9.
Hence, since (

x −1)/(

x + 1) is an increasing function of x on the interval
(1, ∞).
_
κ
2
(A) −1
_
κ
2
(A) + 1


11 −3

11 + 3
≈ .05.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
18 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Therefore (2.15) would predict a reduction in the size of the A-norm error by
a factor of 10
−3
when
2 .05
k
< 10
−3
or when
k > −log
10
(2000)/ log
10
(.05) ≈ 3.3/1.3 ≈ 2.6,
which also predicts termination within three iterations.
We may have more precise information than a single interval containing
σ(A). When we do, the estimate in (2.15) can be very pessimistic. If the
eigenvalues cluster in a small number of intervals, the condition number can
be quite large, but CG can perform very well. We will illustrate this with an
example. Exercise 2.8.5 also covers this point.
Assume that x
0
= 0 and the eigenvalues of A lie in the two intervals (1, 1.5)
and (399, 400). Based on this information the best estimate of the condition
number of A is κ
2
(A) ≤ 400, which, when inserted into (2.15) gives
|x
k
−x

|
A
|x

|
A
≤ 2 (19/21)
k
≈ 2 (.91)
k
.
This would indicate fairly slow convergence. However, if we use as a residual
polynomial ¯ p
3k
∈ T
3k
¯ p
3k
(z) =
(1.25 −z)
k
(400 −z)
2k
(1.25)
k
400
2k
.
It is easy to see that
max
z∈σ(A)
[ ¯ p
3k
(z)[ ≤ (.25/1.25)
k
= (.2)
k
,
which is a sharper estimate on convergence. In fact, (2.15) would predict that
|x
k
−x

|
A
≤ 10
−3
|x

|
A
,
when 2 (.91)
k
< 10
−3
or when
k > −log
10
(2000)/ log
10
(.91) ≈ 3.3/.04 = 82.5.
The estimate based on the clustering gives convergence in 3k iterations when
(.2)
k
≤ 10
−3
or when
k > −3/ log
10
(.2) = 4.3.
Hence (2.15) predicts 83 iterations and the clustering analysis 15 (the smallest
integer multiple of 3 larger than 3 4.3 = 12.9).
From the results above one can see that if the condition number of A is near
one, the CG iteration will converge very rapidly. Even if the condition number
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 19
is large, the iteration will perform well if the eigenvalues are clustered in a few
small intervals. The transformation of the problem into one with eigenvalues
clustered near one (i.e., easier to solve) is called preconditioning. We used
this term before in the context of Richardson iteration and accomplished the
goal by multiplying A by an approximate inverse. In the context of CG, such
a simple approach can destroy the symmetry of the coefficient matrix and a
more subtle implementation is required. We discuss this in ¸ 2.5.
2.4. Implementation
The implementation of CG depends on the amazing fact that once x
k
has been
determined, either x
k
= x

or a search direction p
k+1
,= 0 can be found very
cheaply so that x
k+1
= x
k
+ α
k+1
p
k+1
for some scalar α
k+1
. Once p
k+1
has
been found, α
k+1
is easy to compute from the minimization property of the
iteration. In fact
dφ(x
k
+αp
k+1
)

= 0 (2.16)
for the correct choice of α = α
k+1
. Equation (2.16) can be written as
p
T
k+1
Ax
k
+αp
T
k+1
Ap
k+1
−p
T
k+1
b = 0
leading to
α
k+1
=
p
T
k+1
(b −Ax
k
)
p
T
k+1
Ap
k+1
=
p
T
k+1
r
k
p
T
k+1
Ap
k+1
. (2.17)
If x
k
= x
k+1
then the above analysis implies that α = 0. We show that
this only happens if x
k
is the solution.
Lemma 2.4.1. Let A be spd and let ¦x
k
¦ be the conjugate gradient iterates.
Then
r
T
k
r
l
= 0 for all 0 ≤ l < k. (2.18)
Proof. Since x
k
minimizes φ on x
0
+/
k
, we have, for any ξ ∈ /
k
dφ(x
k
+tξ)
dt
= ∇φ(x
k
+tξ)
T
ξ = 0
at t = 0. Recalling that
∇φ(x) = Ax −b = −r
we have
∇φ(x
k
)
T
ξ = −r
T
k
ξ = 0 for all ξ ∈ /
k
. (2.19)
Since r
l
∈ /
k
for all l < k (see Exercise 2.8.1), this proves (2.18).
Now, if x
k
= x
k+1
, then r
k
= r
k+1
. Lemma 2.4.1 then implies that
|r
k
|
2
2
= r
T
k
r
k
= r
T
k
r
k+1
= 0 and hence x
k
= x

.
The next lemma characterizes the search direction and, as a side effect,
proves that (if we define p
0
= 0) p
T
l
r
k
= 0 for all 0 ≤ l < k ≤ n, unless the
iteration terminates prematurely.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
20 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Lemma 2.4.2. Let A be spd and let ¦x
k
¦ be the conjugate gradient iterates.
If x
k
,= x

then x
k+1
= x
k
+ α
k+1
p
k+1
and p
k+1
is determined up to a scalar
multiple by the conditions
p
k+1
∈ /
k+1
, p
T
k+1
Aξ = 0 for all ξ ∈ /
k
. (2.20)
Proof. Since /
k
⊂ /
k+1
,
∇φ(x
k+1
)
T
ξ = (Ax
k

k+1
Ap
k+1
−b)
T
ξ = 0 (2.21)
for all ξ ∈ /
k
. (2.19) and (2.21) then imply that for all ξ ∈ /
k
,
α
k+1
p
T
k+1
Aξ = −(Ax
k
−b)
T
ξ = −∇φ(x
k
)
T
ξ = 0. (2.22)
This uniquely specifies the direction of p
k+1
as (2.22) implies that p
k+1
∈ /
k+1
is A-orthogonal (i.e., in the scalar product (x, y) = x
T
Ay) to /
k
, a subspace
of dimension one less than /
k+1
.
The condition p
T
k+1
Aξ = 0 is called A-conjugacy of p
k+1
to /
k
. Now, any
p
k+1
satisfying (2.20) can, up to a scalar multiple, be expressed as
p
k+1
= r
k
+w
k
with w
k
∈ /
k
. While one might think that w
k
would be hard to compute, it
is, in fact, trivial. We have the following theorem.
Theorem 2.4.1. Let A be spd and assume that r
k
,= 0. Define p
0
= 0.
Then
p
k+1
= r
k

k+1
p
k
for some β
k+1
and k ≥ 0. (2.23)
Proof. By Lemma 2.4.2 and the fact that /
k
= span(r
0
, . . . , r
k−1
), we need
only verify that a β
k+1
can be found so that if p
k+1
is given by (2.23) then
p
T
k+1
Ar
l
= 0
for all 0 ≤ l ≤ k −1.
Let p
k+1
be given by (2.23). Then for any l ≤ k
p
T
k+1
Ar
l
= r
T
k
Ar
l

k+1
p
T
k
Ar
l
.
If l ≤ k −2, then r
l
∈ /
l+1
⊂ /
k−1
. Lemma 2.4.2 then implies that
p
T
k+1
Ar
l
= 0 for 0 ≤ l ≤ k −2.
It only remains to solve for β
k+1
so that p
T
k+1
Ar
k−1
= 0. Trivially
β
k+1
= −r
T
k
Ar
k−1
/p
T
k
Ar
k−1
(2.24)
provided p
T
k
Ar
k−1
,= 0. Since
r
k
= r
k−1
−α
k
Ap
k
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 21
we have
r
T
k
r
k−1
= |r
k−1
|
2
2
−α
k
p
T
k
Ar
k−1
.
Since r
T
k
r
k−1
= 0 by Lemma 2.4.1 we have
p
T
k
Ar
k−1
= |r
k−1
|
2
2

k
,= 0. (2.25)
This completes the proof.
The common implementation of conjugate gradient uses a different form
for α
k
and β
k
than given in (2.17) and (2.24).
Lemma 2.4.3. Let A be spd and assume that r
k
,= 0. Then
α
k+1
=
|r
k
|
2
2
p
T
k+1
Ap
k+1
(2.26)
and
β
k+1
=
|r
k
|
2
2
|r
k−1
|
2
2
. (2.27)
Proof. Note that for k ≥ 0
p
T
k+1
r
k+1
= r
T
k
r
k+1

k+1
p
T
k
r
k+1
= 0 (2.28)
by Lemma 2.4.2. An immediate consequence of (2.28) is that p
T
k
r
k
= 0 and
hence
p
T
k+1
r
k
= (r
k

k+1
p
k
)
T
r
k
= |r
k
|
2
2
. (2.29)
Taking scalar products of both sides of
r
k+1
= r
k
−α
k+1
Ap
k+1
with p
k+1
and using (2.29) gives
0 = p
T
k+1
r
k
−α
k+1
p
T
k+1
Ap
k+1
= |r
T
k
|
2
2
−α
k+1
p
T
k+1
Ap
k+1
,
which is equivalent to (2.26).
To get (2.27) note that p
T
k+1
Ap
k
= 0 and hence (2.23) implies that
β
k+1
=
−r
T
k
Ap
k
p
T
k
Ap
k
. (2.30)
Also note that
p
T
k
Ap
k
= p
T
k
A(r
k−1

k
p
k−1
)
= p
T
k
Ar
k−1

k
p
T
k
Ap
k−1
= p
T
k
Ar
k−1
.
(2.31)
Now combine (2.30), (2.31), and (2.25) to get
β
k+1
=
−r
T
k
Ap
k
α
k
|r
k−1
|
2
2
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
22 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Now take scalar products of both sides of
r
k
= r
k−1
−α
k
Ap
k
with r
k
and use Lemma 2.4.1 to get
|r
k
|
2
2
= −α
k
r
T
k
Ap
k
.
Hence (2.27) holds.
The usual implementation reflects all of the above results. The goal is to
find, for a given , a vector x so that |b−Ax|
2
≤ |b|
2
. The input is the initial
iterate x, which is overwritten with the solution, the right hand side b, and a
routine which computes the action of A on a vector. We limit the number of
iterations to kmax and return the solution, which overwrites the initial iterate
x and the residual norm.
Algorithm 2.4.1. cg(x, b, A, , kmax)
1. r = b −Ax, ρ
0
= |r|
2
2
, k = 1.
2. Do While

ρ
k−1
> |b|
2
and k < kmax
(a) if k = 1 then p = r
else
β = ρ
k−1

k−2
and p = r +βp
(b) w = Ap
(c) α = ρ
k−1
/p
T
w
(d) x = x +αp
(e) r = r −αw
(f) ρ
k
= |r|
2
2
(g) k = k + 1
Note that the matrix A itself need not be formed or stored, only a routine
for matrix-vector products is required. Krylov space methods are often called
matrix-free for that reason.
Now, consider the costs. We need store only the four vectors x, w, p, and r.
Each iteration requires a single matrix-vector product (to compute w = Ap),
two scalar products (one for p
T
w and one to compute ρ
k
= |r|
2
2
), and three
operations of the form ax +y, where x and y are vectors and a is a scalar.
It is remarkable that the iteration can progress without storing a basis for
the entire Krylov subspace. As we will see in the section on GMRES, this is
not the case in general. The spd structure buys quite a lot.
2.5. Preconditioning
To reduce the condition number, and hence improve the performance of the
iteration, one might try to replace Ax = b by another spd system with the
same solution. If M is a spd matrix that is close to A
−1
, then the eigenvalues
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 23
of MA will be clustered near one. However MA is unlikely to be spd, and
hence CG cannot be applied to the system MAx = Mb.
In theory one avoids this difficulty by expressing the preconditioned
problem in terms of B, where B is spd, A = B
2
, and by using a two-sided
preconditioner, S ≈ B
−1
(so M = S
2
). Then the matrix SAS is spd and its
eigenvalues are clustered near one. Moreover the preconditioned system
SASy = Sb
has y

= S
−1
x

as a solution, where Ax

= b. Hence x

can be recovered
from y

by multiplication by S. One might think, therefore, that computing
S (or a subroutine for its action on a vector) would be necessary and that a
matrix-vector multiply by SAS would incur a cost of one multiplication by A
and two by S. Fortunately, this is not the case.
If y
k
, ˆ r
k
, ˆ p
k
are the iterate, residual, and search direction for CG applied
to SAS and we let
x
k
= Sˆ y
k
, r
k
= S
−1
ˆ r
k
, p
k
= Sˆ p
k
, and z
k
= Sˆ r
k
,
then one can perform the iteration directly in terms of x
k
, A, and M. The
reader should verify that the following algorithm does exactly that. The input
is the same as that for Algorithm cg and the routine to compute the action of
the preconditioner on a vector. Aside from the preconditioner, the arguments
to pcg are the same as those to Algorithm cg.
Algorithm 2.5.1. pcg(x, b, A, M, , kmax)
1. r = b −Ax, ρ
0
= |r|
2
2
, k = 1
2. Do While

ρ
k−1
> |b|
2
and k < kmax
(a) z = Mr
(b) τ
k−1
= z
T
r
(c) if k = 1 then β = 0 and p = z
else
β = τ
k−1

k−2
, p = z +βp
(d) w = Ap
(e) α = τ
k−1
/p
T
w
(f) x = x +αp
(g) r = r −αw
(h) ρ
k
= r
T
r
(i) k = k + 1
Note that the cost is identical to CG with the addition of
• the application of the preconditioner M in step 2a and
• the additional inner product required to compute τ
k
in step 2b.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
24 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Of these costs, the application of the preconditioner is usually the larger. In
the remainder of this section we briefly mention some classes of preconditioners.
A more complete and detailed discussion of preconditioners is in [8] and a
concise survey with many pointers to the literature is in [12].
Some effective preconditioners are based on deep insight into the structure
of the problem. See [124] for an example in the context of partial differential
equations, where it is shown that certain discretized second-order elliptic
problems on simple geometries can be very well preconditioned with fast
Poisson solvers [99], [188], and [187]. Similar performance can be obtained from
multigrid [99], domain decomposition, [38], [39], [40], and alternating direction
preconditioners [8], [149], [193], [194]. We use a Poisson solver preconditioner
in the examples in ¸ 2.7 and ¸ 3.7 as well as for nonlinear problems in ¸ 6.4.2
and ¸ 8.4.2.
One commonly used and easily implemented preconditioner is Jacobi
preconditioning, where M is the inverse of the diagonal part of A. One can also
use other preconditioners based on the classical stationary iterative methods,
such as the symmetric Gauss–Seidel preconditioner (1.18). For applications to
partial differential equations, these preconditioners may be somewhat useful,
but should not be expected to have dramatic effects.
Another approach is to apply a sparse Cholesky factorization to the
matrix A (thereby giving up a fully matrix-free formulation) and discarding
small elements of the factors and/or allowing only a fixed amount of storage
for the factors. Such preconditioners are called incomplete factorization
preconditioners. So if A = LL
T
+ E, where E is small, the preconditioner
is (LL
T
)
−1
and its action on a vector is done by two sparse triangular solves.
We refer the reader to [8], [127], and [44] for more detail.
One could also attempt to estimate the spectrum of A, find a polynomial
p such that 1 −zp(z) is small on the approximate spectrum, and use p(A) as a
preconditioner. This is called polynomial preconditioning. The preconditioned
system is
p(A)Ax = p(A)b
and we would expect the spectrum of p(A)A to be more clustered near z = 1
than that of A. If an interval containing the spectrum can be found, the
residual polynomial q(z) = 1 − zp(z) of smallest L

norm on that interval
can be expressed in terms of Chebyshev [161] polynomials. Alternatively
q can be selected to solve a least squares minimization problem [5], [163].
The preconditioning p can be directly recovered from q and convergence rate
estimates made. This technique is used to prove the estimate (2.15), for
example. The cost of such a preconditioner, if a polynomial of degree K is
used, is K matrix-vector products for each application of the preconditioner
[5]. The performance gains can be very significant and the implementation is
matrix-free.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 25
2.6. CGNR and CGNE
If A is nonsingular and nonsymmetric, one might consider solving Ax = b by
applying CG to the normal equations
A
T
Ax = A
T
b. (2.32)
This approach [103] is called CGNR [71], [78], [134]. The reason for this name
is that the minimization property of CG as applied to (2.32) asserts that
|x

−x|
2
A
T
A
= (x

−x)
T
A
T
A(x

−x)
= (Ax

−Ax)
T
(Ax

−Ax) = (b −Ax)
T
(b −Ax) = |r|
2
is minimized over x
0
+/
k
at each iterate. Hence the name Conjugate Gradient
on the Normal equations to minimize the Residual.
Alternatively, one could solve
AA
T
y = b (2.33)
and then set x = A
T
y to solve Ax = b. This approach [46] is now called CGNE
[78], [134]. The reason for this name is that the minimization property of CG
as applied to (2.33) asserts that if y

is the solution to (2.33) then
|y

−y|
2
AA
T
= (y

−y)
T
(AA
T
)(y

−y) = (A
T
y

−A
T
y)
T
(A
T
y

−A
T
y)
= |x

−x|
2
2
is minimized over y
0
+/
k
at each iterate. Conjugate Gradient on the Normal
equations to minimize the Error.
The advantages of this approach are that all the theory for CG carries over
and the simple implementation for both CG and PCG can be used. There
are three disadvantages that may or may not be serious. The first is that the
condition number of the coefficient matrix A
T
A is the square of that of A.
The second is that two matrix-vector products are needed for each CG iterate
since w = A
T
Ap = A
T
(Ap) in CGNR and w = AA
T
p = A(A
T
p) in CGNE.
The third, more important, disadvantage is that one must compute the action
of A
T
on a vector as part of the matrix-vector product involving A
T
A. As we
will see in the chapter on nonlinear problems, there are situations where this
is not possible.
The analysis with residual polynomials is similar to that for CG. We
consider the case for CGNR, the analysis for CGNE is essentially the same.
As above, when we consider the A
T
A norm of the error, we have
|x

−x|
2
A
T
A
= (x

−x)
T
A
T
A(x

−x) = |A(x

−x)|
2
2
= |r|
2
2
.
Hence, for any residual polynomial ¯ p
k
∈ T
k
,
|r
k
|
2
≤ |¯ p
k
(A
T
A)r
0
|
2
≤ |r
0
|
2
max
z∈σ(A
T
A)
[ ¯ p
k
(z)[. (2.34)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
26 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
There are two major differences between (2.34) and (2.7). The estimate is
in terms of the l
2
norm of the residual, which corresponds exactly to the
termination criterion, hence we need not prove a result like Lemma 2.3.2. Most
significantly, the residual polynomial is to be maximized over the eigenvalues
of A
T
A, which is the set of the squares of the singular values of A. Hence the
performance of CGNR and CGNE is determined by the distribution of singular
values.
2.7. Examples for preconditioned conjugate iteration
In the collection of MATLAB codes we provide a code for preconditioned
conjugate gradient iteration. The inputs, described in the comment lines,
are the initial iterate, x
0
, the right hand side vector b, MATLAB functions for
the matrix-vector product and (optionally) the preconditioner, and iteration
parameters to specify the maximum number of iterations and the termination
criterion. On return the code supplies the approximate solution x and the
history of the iteration as the vector of residual norms.
We consider the discretization of the partial differential equation
−∇
˙
(a(x, y)∇u) = f(x, y) (2.35)
on 0 < x, y < 1 subject to homogeneous Dirichlet boundary conditions
u(x, 0) = u(x, 1) = u(0, y) = u(1, y) = 0, 0 < x, y < 1.
One can verify [105] that the differential operator is positive definite in the
Hilbert space sense and that the five-point discretization described below is
positive definite if a > 0 for all 0 ≤ x, y ≤ 1 (Exercise 2.8.10).
We discretize with a five-point centered difference scheme with n
2
points
and mesh width h = 1/(n + 1). The unknowns are
u
ij
≈ u(x
i
, x
j
)
where x
i
= ih for 1 ≤ i ≤ n. We set
u
0j
= u
(n+1)j
= u
i0
= u
i(n+1)
= 0,
to reflect the boundary conditions, and define
α
ij
= −a(x
i
, x
j
)h
−2
/2.
We express the discrete matrix-vector product as
(Au)
ij
= (α
ij

(i+1)j
)(u
(i+1)j
−u
ij
)
−(α
(i−1)j

ij
)(u
ij
−u
(i−1)j
) + (α
i(j+1)

ij
)(u
i(j+1)
−u
ij
)
−(α
ij

i(j−1)
)(u
ij
−u
i(j−1)
)
(2.36)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 27
for 1 ≤ i, j ≤ n.
For the MATLAB implementation we convert freely from the representa-
tion of u as a two-dimensional array (with the boundary conditions added),
which is useful for computing the action of A on u and applying fast solvers,
and the representation as a one-dimensional array, which is what pcgsol ex-
pects to see. See the routine fish2d in collection of MATLAB codes for an
example of how to do this in MATLAB.
For the computations reported in this section we took a(x, y) = cos(x) and
took the right hand side so that the exact solution was the discretization of
10xy(1 −x)(1 −y) exp(x
4.5
).
The initial iterate was u
0
= 0.
In the results reported here we took n = 31 resulting in a system with
N = n
2
= 961 unknowns. We expect second-order accuracy from the method
and accordingly we set termination parameter = h
2
= 1/1024. We allowed
up to 100 CG iterations. The initial iterate was the zero vector. We will report
our results graphically, plotting |r
k
|
2
/|b|
2
on a semi-log scale.
In Fig. 2.1 the solid line is a plot of |r
k
|/|b| and the dashed line a plot of
|u

−u
k
|
A
/|u

−u
0
|
A
. Note that the reduction in |r| is not monotone. This is
consistent with the theory, which predicts decrease in |e|
A
but not necessarily
in |r| as the iteration progresses. Note that the unpreconditioned iteration is
slowly convergent. This can be explained by the fact that the eigenvalues are
not clustered and
κ(A) = O(1/h
2
) = O(n
2
) = O(N)
and hence (2.15) indicates that convergence will be slow. The reader is asked
to quantify this in terms of execution times in Exercise 2.8.9. This example
illustrates the importance of a good preconditioner. Even the unpreconditioned
iteration, however, is more efficient that the classical stationary iterative
methods.
For a preconditioner we use a Poisson solver. By this we mean an operator
G such that v = Gw is the solution of the discrete form of
−v
xx
−v
yy
= w,
subject to homogeneous Dirichlet boundary conditions. The effectiveness of
such a preconditioner has been analyzed in [124] and some of the many ways
to implement the solver efficiently are discussed in [99], [188], [186], and [187].
The properties of CG on the preconditioned problem in the continuous
case have been analyzed in [48]. For many types of domains and boundary
conditions, Poisson solvers can be designed to take advantage of vector and/or
parallel architectures or, in the case of the MATLAB environment used in
this book, designed to take advantage of fast MATLAB built-in functions.
Because of this their execution time is less than a simple count of floating-
point operations would indicate. The fast Poisson solver in the collection of
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
28 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
0 10 20 30 40 50 60
10
-4
10
-3
10
-2
10
-1
10
0
10
1
Iterations
R
e
l
a
t
i
v
e

R
e
s
i
d
u
a
l

a
n
d

A
-
n
o
r
m

o
f

E
r
r
o
r
Fig. 2.1. CG for 2-D elliptic equation.
0 10 20 30 40 50 60
10
-4
10
-3
10
-2
10
-1
10
0
10
1
Iterations
R
e
l
a
t
i
v
e

R
e
s
i
d
u
a
l
Fig. 2.2. PCG for 2-D elliptic equation.
codes fish2d is based on the MATLAB fast Fourier transform, the built-in
function fft.
In Fig. 2.2 the solid line is the graph of |r
k
|
2
/|b|
2
for the preconditioned
iteration and the dashed line for the unpreconditioned. The preconditioned
iteration required 5 iterations for convergence and the unpreconditioned
iteration 52. Not only does the preconditioned iteration converge more
rapidly, but the number of iterations required to reduce the relative residual
by a given amount is independent of the mesh spacing [124]. We caution
the reader that the preconditioned iteration is not as much faster than the
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 29
unpreconditioned one as the iteration count would suggest. The MATLAB
flops command indicates that the unpreconditioned iteration required roughly
1.2 million floating-point operations while the preconditioned iteration required
.87 million floating-point operations. Hence, the cost of the preconditioner is
considerable. In the MATLAB environment we used, the execution time of
the preconditioned iteration was about 60% of that of the unpreconditioned.
As we remarked above, this speed is a result of the efficiency of the MATLAB
fast Fourier transform. In Exercise 2.8.11 you are asked to compare execution
times for your own environment.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
30 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
2.8. Exercises on conjugate gradient
2.8.1. Let ¦x
k
¦ be the conjugate gradient iterates. Prove that r
l
∈ /
k
for all
l < k.
2.8.2. Let A be spd. Show that there is a spd B such that B
2
= A. Is B
unique?
2.8.3. Let Λ be a diagonal matrix with Λ
ii
= λ
i
and let p be a polynomial.
Prove that |p(Λ)| = max
i
[p(λ
i
)[ where | | is any induced matrix norm.
2.8.4. Prove Theorem 2.2.3.
2.8.5. Assume that A is spd and that
σ(A) ⊂ (1, 1.1) ∪ (2, 2.2).
Give upper estimates based on (2.6) for the number of CG iterations
required to reduce the A norm of the error by a factor of 10
−3
and for
the number of CG iterations required to reduce the residual by a factor
of 10
−3
.
2.8.6. For the matrix A in problem 5, assume that the cost of a matrix vector
multiply is 4N floating-point multiplies. Estimate the number of floating-
point operations reduce the A norm of the error by a factor of 10
−3
using
CG iteration.
2.8.7. Let A be a nonsingular matrix with all singular values in the interval
(1, 2). Estimate the number of CGNR/CGNE iterations required to
reduce the relative residual by a factor of 10
−4
.
2.8.8. Show that if A has constant diagonal then PCG with Jacobi precondi-
tioning produces the same iterates as CG with no preconditioning.
2.8.9. Assume that A is N N, nonsingular, and spd. If κ(A) = O(N), give
a rough estimate of the number of CG iterates required to reduce the
relative residual to O(1/N).
2.8.10. Prove that the linear transformation given by (2.36) is symmetric and
positive definite on R
n
2
if a(x, y) > 0 for all 0 ≤ x, y ≤ 1.
2.8.11. Duplicate the results in ¸ 2.7 for example, in MATLAB by writing the
matrix-vector product routines and using the MATLAB codes pcgsol
and fish2d. What happens as N is increased? How are the performance
and accuracy affected by changes in a(x, y)? Try a(x, y) =

.1 +x and
examine the accuracy of the result. Explain your findings. Compare
the execution times on your computing environment (using the cputime
command in MATLAB, for instance).
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
CONJUGATE GRADIENT ITERATION 31
2.8.12. Use the Jacobi and symmetric Gauss–Seidel iterations from Chapter 1
to solve the elliptic boundary value problem in ¸ 2.7. How does the
performance compare to CG and PCG?
2.8.13. Implement Jacobi (1.17) and symmetric Gauss–Seidel (1.18) precondi-
tioners for the elliptic boundary value problem in ¸ 2.7. Compare the
performance with respect to both computer time and number of itera-
tions to preconditioning with the Poisson solver.
2.8.14. Modify pcgsol so that φ(x) is computed and stored at each iterate
and returned on output. Plot φ(x
n
) as a function of n for each of the
examples.
2.8.15. Apply CG and PCG to solve the five-point discretization of
−u
xx
(x, y) −u
yy
(x, y) +e
x+y
u(x, y) = 1, 0 < x, y, < 1,
subject to the inhomogeneous Dirichlet boundary conditions
u(x, 0) = u(x, 1) = u(1, y) = 0, u(0, y) = 1, 0 < x, y < 1.
Experiment with different mesh sizes and preconditioners (Fast Poisson
solver, Jacobi, and symmetric Gauss–Seidel).
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
32 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 3
GMRES Iteration
3.1. The minimization property and its consequences
The GMRES (Generalized Minimum RESidual) was proposed in 1986 in [167]
as a Krylov subspace method for nonsymmetric systems. Unlike CGNR,
GMRES does not require computation of the action of A
T
on a vector. This is
a significant advantage in many cases. The use of residual polynomials is made
more complicated because we cannot use the spectral theorem to decompose
A. Moreover, one must store a basis for /
k
, and therefore storage requirements
increase as the iteration progresses.
The kth (k ≥ 1) iteration of GMRES is the solution to the least squares
problem
minimize
x∈x
0
+K
k
|b −Ax|
2
. (3.1)
The beginning of this section is much like the analysis for CG. Note that
if x ∈ x
0
+/
k
then
x = x
0
+
k−1

j=0
γ
j
A
j
r
0
and so
b −Ax = b −Ax
0

k−1

j=0
γ
j
A
j+1
r
0
= r
0

k

j=1
γ
j−1
A
j
r
0
.
Hence if x ∈ x
0
+/
k
then r = ¯ p(A)r
0
where ¯ p ∈ T
k
is a residual polynomial.
We have just proved the following result.
Theorem 3.1.1. Let A be nonsingular and let x
k
be the kth GMRES
iteration. Then for all ¯ p
k
∈ T
k
|r
k
|
2
= min
p∈P
k
|¯ p(A)r
0
|
2
≤ |¯ p
k
(A)r
0
|
2
. (3.2)
From this we have the following corollary.
Corollary 3.1.1. Let A be nonsingular and let x
k
be the kth GMRES
iteration. Then for all ¯ p
k
∈ T
k
|r
k
|
2
|r
0
|
2
≤ |¯ p
k
(A)|
2
. (3.3)
33
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
34 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
We can apply the corollary to prove finite termination of the GMRES
iteration.
Theorem 3.1.2. Let A be nonsingular. Then the GMRES algorithm will
find the solution within N iterations.
Proof. The characteristic polynomial of A is p(z) = det(A − zI). p has
degree N, p(0) = det(A) ,= 0 since A is nonsingular, and so
¯ p
N
(z) = p(z)/p(0) ∈ T
N
is a residual polynomial. It is well known [141] that p(A) = ¯ p
N
(A) = 0. By
(3.3), r
N
= b −Ax
N
= 0 and hence x
N
is the solution.
In Chapter 2 we applied the spectral theorem to obtain more precise infor-
mation on convergence rates. This is not an option for general nonsymmetric
matrices. However, if A is diagonalizable we may use (3.2) to get information
from clustering of the spectrum just like we did for CG. We pay a price in that
we must use complex arithmetic for the only time in this book. Recall that
A is diagonalizable if there is a nonsingular (possibly complex!) matrix V such
that
A = V ΛV
−1
.
Here Λ is a (possibly complex!) diagonal matrix with the eigenvalues of A on
the diagonal. If A is a diagonalizable matrix and p is a polynomial then
p(A) = V p(Λ)V
−1
A is normal if the diagonalizing transformation V is orthogonal. In that case
the columns of V are the eigenvectors of A and V
−1
= V
H
. Here V
H
is the
complex conjugate transpose of V . In the remainder of this section we must
use complex arithmetic to analyze the convergence. Hence we will switch to
complex matrices and vectors. Recall that the scalar product in C
N
, the space
of complex N-vectors, is x
H
y. In particular, we will use the l
2
norm in C
N
.
Our use of complex arithmetic will be implicit for the most part and is needed
only so that we may admit the possibility of complex eigenvalues of A.
We can use the structure of a diagonalizable matrix to prove the following
result.
Theorem 3.1.3. Let A = V ΛV
−1
be a nonsingular diagonalizable matrix.
Let x
k
be the kth GMRES iterate. Then for all ¯ p
k
∈ T
k
|r
k
|
2
|r
0
|
2
≤ κ
2
(V ) max
z∈σ(A)
[ ¯ p
k
(z)[. (3.4)
Proof. Let ¯ p
k
∈ T
k
. We can easily estimate |¯ p
k
(A)|
2
by
|¯ p
k
(A)|
2
≤ |V |
2
|V
−1
|
2
|¯ p
k
(Λ)|
2
≤ κ
2
(V ) max
z∈σ(A)
[ ¯ p
k
(z)[,
as asserted.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 35
It is not clear how one should estimate the condition number of the
diagonalizing transformation if it exists. If A is normal, of course, κ
2
(V ) = 1.
As we did for CG, we look at some easy consequences of (3.3) and (3.4).
Theorem 3.1.4. Let A be a nonsingular diagonalizable matrix. Assume
that A has only k distinct eigenvalues. Then GMRES will terminate in at most
k iterations.
Theorem 3.1.5. Let A be a nonsingular normal matrix. Let b be a linear
combination of k of the eigenvectors of A
b =
k

l=1
γ
l
u
i
l
.
Then the GMRES iteration, with x
0
= 0, for Ax = b will terminate in at most
k iterations.
3.2. Termination
As is the case with CG, GMRES is best thought of as an iterative method.
The convergence rate estimates for the diagonalizable case will involve κ
2
(V ),
but will otherwise resemble those for CG. If A is not diagonalizable, rate
estimates have been derived in [139], [134], [192], [33], and [34]. As the set of
nondiagonalizable matrices has measure zero in the space of N N matrices,
the chances are very high that a computed matrix will be diagonalizable. This
is particularly so for the finite difference Jacobian matrices we consider in
Chapters 6 and 8. Hence we confine our attention to diagonalizable matrices.
As was the case with CG, we terminate the iteration when
|r
k
|
2
≤ η|b|
2
(3.5)
for the purposes of this example. We can use (3.3) and (3.4) directly to estimate
the first k such that (3.5) holds without requiring a lemma like Lemma 2.3.2.
Again we look at examples. Assume that A = V ΛV
−1
is diagonalizable,
that the eigenvalues of A lie in the interval (9, 11), and that κ
2
(V ) = 100.
We assume that x
0
= 0 and hence r
0
= b. Using the residual polynomial
¯ p
k
(z) = (10 −z)
k
/10
k
we find
|r
k
|
2
|r
0
|
2
≤ (100)10
−k
= 10
2−k
.
Hence (3.5) holds when 10
2−k
< η or when
k > 2 + log
10
(η).
Assume that |I − A|
2
≤ ρ < 1. Let ¯ p
k
(z) = (1 − z)
k
. It is a direct
consequence of (3.2) that
|r
k
|
2
≤ ρ
k
|r
0
|
2
. (3.6)
The estimate (3.6) illustrates the potential benefits of a good approximate
inverse preconditioner.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
36 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
The convergence estimates for GMRES in the nonnormal case are much
less satisfying that those for CG, CGNR, CGNE, or GMRES in the normal
case. This is a very active area of research and we refer to [134], [33], [120],
[34], and [36] for discussions of and pointers to additional references to several
questions related to nonnormal matrices.
3.3. Preconditioning
Preconditioning for GMRES and other methods for nonsymmetric problems is
different from that for CG. There is no concern that the preconditioned system
be spd and hence (3.6) essentially tells the whole story. However there are two
different ways to view preconditioning. If one can find M such that
|I −MA|
2
= ρ < 1,
then applying GMRES to MAx = Mb allows one to apply (3.6) to the
preconditioned system. Preconditioning done in this way is called left
preconditioning. If r = MAx − Mb is the residual for the preconditioned
system, we have, if the product MA can be formed without error,
|e
k
|
2
|e
0
|
2
≤ κ
2
(MA)
|r
k
|
2
|r
0
|
2
,
by Lemma 1.1.1. Hence, if MA has a smaller condition number than A, we
might expect the relative residual of the preconditioned system to be a better
indicator of the relative error than the relative residual of the original system.
If
|I −AM|
2
= ρ < 1,
one can solve the system AMy = b with GMRES and then set x = My. This is
called right preconditioning. The residual of the preconditioned problem is the
same as that of the unpreconditioned problem. Hence, the value of the relative
residuals as estimators of the relative error is unchanged. Right preconditioning
has been used as the basis for a method that changes the preconditioner as the
iteration progresses [166].
One important aspect of implementation is that, unlike PCG, one can
apply the algorithm directly to the system MAx = Mb (or AMy = b). Hence,
one can write a single matrix-vector product routine for MA (or AM) that
includes both the application of A to a vector and that of the preconditioner.
Most of the preconditioning ideas mentioned in ¸ 2.5 are useful for GMRES
as well. In the examples in ¸ 3.7 we use the Poisson solver preconditioner for
nonsymmetric partial differential equations. Multigrid [99] and alternating
direction [8], [182] methods have similar performance and may be more
generally applicable. Incomplete factorization (LU in this case) preconditioners
can be used [165] as can polynomial preconditioners. Some hybrid algorithms
use the GMRES/Arnoldi process itself to construct polynomial preconditioners
for GMRES or for Richardson iteration [135], [72], [164], [183]. Again we
mention [8] and [12] as a good general references for preconditioning.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 37
3.4. GMRES implementation: Basic ideas
Recall that the least squares problem defining the kth GMRES iterate is
minimize
x∈x
0
+K
k
|b −Ax|
2
.
Suppose one had an orthogonal projector V
k
onto /
k
. Then any z ∈ /
k
can
be written as
z =
k

l=1
y
l
v
k
l
,
where v
k
l
is the lth column of V
k
. Hence we can convert (3.1) to a least squares
problem in R
k
for the coefficient vector y of z = x −x
0
. Since
x −x
0
= V
k
y
for some y ∈ R
k
, we must have x
k
= x
0
+V
k
y where y minimizes
|b −A(x
0
+V
k
y)|
2
= |r
0
−AV
k
y|
2
.
Hence, our least squares problem in R
k
is
minimize
y∈R
k|r
0
−AV
k
y|
2
. (3.7)
This is a standard linear least squares problem that could be solved by a QR
factorization, say. The problem with such a direct approach is that the matrix
vector product of A with the basis matrix V
k
must be taken at each iteration.
If one uses Gram–Schmidt orthogonalization, however, one can represent
(3.7) very efficiently and the resulting least squares problem requires no extra
multiplications of A with vectors. The Gram–Schmidt procedure for formation
of an orthonormal basis for /
k
is called the Arnoldi [4] process. The data are
vectors x
0
and b, a map that computes the action of A on a vector, and a
dimension k. The algorithm computes an orthonormal basis for /
k
and stores
it in the columns of V .
Algorithm 3.4.1. arnoldi(x
0
, b, A, k, V )
1. Define r
0
= b −Ax
0
and v
1
= r
0
/|r
0
|
2
.
2. For i = 1, . . . , k −1
v
i+1
=
Av
i

i
j=1
((Av
i
)
T
v
j
)v
j
|Av
i

i
j=1
((Av
i
)
T
v
j
)v
j
|
2
If there is never a division by zero in step 2 of Algorithm arnoldi, then
the columns of the matrix V
k
are an orthonormal basis for /
k
. A division by
zero is referred to as breakdown and happens only if the solution to Ax = b is
in x
0
+/
k−1
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
38 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Lemma 3.4.1. Let A be nonsingular, let the vectors v
j
be generated by
Algorithm arnoldi, and let i be the smallest integer for which
Av
i

i

j=1
((Av
i
)
T
v
j
)v
j
= 0.
Then x = A
−1
b ∈ x
0
+/
i
.
Proof. By hypothesis Av
i
∈ /
i
and hence A/
i
⊂ /
i
. Since the columns of
V
i
are an orthonormal basis for /
i
, we have
AV
i
= V
i
H,
where H is an i i matrix. H is nonsingular since A is. Setting β = |r
0
|
2
and e
1
= (1, 0, . . . , 0)
T
∈ R
i
we have
|r
i
|
2
= |b −Ax
i
|
2
= |r
0
−A(x
i
−x
0
)|
2
.
Since x
i
− x
0
∈ /
i
there is y ∈ R
i
such that x
i
− x
0
= V
i
y. Since r
0
= βV
i
e
1
and V
i
is an orthogonal matrix
|r
i
|
2
= |V
i
(βe
1
−Hy)|
2
= |βe
1
−Hy|
R
i+1,
where | |
R
k+1 denotes the Euclidean norm in R
k+1
.
Setting y = βH
−1
e
1
proves that r
i
= 0 by the minimization property.
The upper Hessenberg structure can be exploited to make the solution of
the least squares problems very efficient [167].
If the Arnoldi process does not break down, we can use it to implement
GMRES in an efficient way. Set h
ji
= (Av
j
)
T
v
i
. By the Gram–Schmidt
construction, the k+1k matrix H
k
whose entries are h
ij
is upper Hessenberg,
i.e., h
ij
= 0 if i > j +1. The Arnoldi process (unless it terminates prematurely
with a solution) produces matrices ¦V
k
¦ with orthonormal columns such that
AV
k
= V
k+1
H
k
. (3.8)
Hence, for some y
k
∈ R
k
,
r
k
= b −Ax
k
= r
0
−A(x
k
−x
0
) = V
k+1
(βe
1
−H
k
y
k
).
Hence x
k
= x
0
+V
k
y
k
, where y
k
minimizes |βe
1
−H
k
y|
2
over R
k
. Note that
when y
k
has been computed, the norm of r
k
can be found without explicitly
forming x
k
and computing r
k
= b −Ax
k
. We have, using the orthogonality of
V
k+1
,
|r
k
|
2
= |V
k+1
(βe
1
−H
k
y
k
)|
2
= |βe
1
−H
k
y
k
|
R
k+1. (3.9)
The goal of the iteration is to find, for a given , a vector x so that
|b −Ax|
2
≤ |b|
2
.
The input is the initial iterate, x, the right-hand side b, and a map that
computes the action of A on a vector. We limit the number of iterations
to kmax and return the solution, which overwrites the initial iterate x and the
residual norm.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 39
Algorithm 3.4.2. gmresa(x, b, A, , kmax, ρ)
1. r = b −Ax, v
1
= r/|r|
2
, ρ = |r|
2
, β = ρ, k = 0
2. While ρ > |b|
2
and k < kmax do
(a) k = k + 1
(b) for j = 1, . . . , k
h
jk
= (Av
k
)
T
v
j
(c) v
k+1
= Av
k

k
j=1
h
jk
v
j
(d) h
k+1,k
= |v
k+1
|
2
(e) v
k+1
= v
k+1
/|v
k+1
|
2
(f) e
1
= (1, 0, . . . , 0)
T
∈ R
k+1
Minimize |βe
1
−H
k
y
k
|
R
k+1 over R
k
to obtain y
k
.
(g) ρ = |βe
1
−H
k
y
k
|
R
k+1.
3. x
k
= x
0
+V
k
y
k
.
Note that x
k
is only computed upon termination and is not needed within
the iteration. It is an important property of GMRES that the basis for the
Krylov space must be stored as the iteration progress. This means that in
order to perform k GMRES iterations one must store k vectors of length N.
For very large problems this becomes prohibitive and the iteration is restarted
when the available room for basis vectors is exhausted. One way to implement
this is to set kmax to the maximum number m of vectors that one can store,
call GMRES and explicitly test the residual b−Ax
k
if k = m upon termination.
If the norm of the residual is larger than , call GMRES again with x
0
= x
k
,
the result from the previous call. This restarted version of the algorithm is
called GMRES(m) in [167]. There is no general convergence theorem for the
restarted algorithm and restarting will slow the convergence down. However,
when it works it can significantly reduce the storage costs of the iteration. We
discuss implementation of GMRES(m) later in this section.
Algorithm gmresa can be implemented very straightforwardly in MAT-
LAB. Step 2f can be done with a single MATLAB command, the backward
division operator, at a cost of O(k
3
) floating-point operations. There are more
efficient ways to solve the least squares problem in step 2f, [167], [197], and we
use the method of [167] in the collection of MATLAB codes. The savings are
slight if k is small relative to N, which is often the case for large problems, and
the simple one-line MATLAB approach can be efficient for such problems.
A more serious problem with the implementation proposed in Algo-
rithm gmresa is that the vectors v
j
may become nonorthogonal as a result of
cancellation errors. If this happens, (3.9), which depends on this orthogonality,
will not hold and the residual and approximate solution could be inaccurate. A
partial remedy is to replace the classical Gram–Schmidt orthogonalization in
Algorithm gmresa with modified Gram–Schmidt orthogonalization. We replace
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
40 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
the loop in step 2c of Algorithm gmresa with
v
k+1
= Av
k
for j = 1, . . . k
v
k+1
= v
k+1
−(v
T
k+1
v
j
)v
j
.
While modified Gram–Schmidt and classical Gram–Schmidt are equivalent in
infinite precision, the modified form is much more likely in practice to maintain
orthogonality of the basis.
We illustrate this point with a simple example from [128], doing the
computations in MATLAB. Let δ = 10
−7
and define
A =
_
_
_
1 1 1
δ δ 0
δ 0 δ
_
_
_.
We orthogonalize the columns of A with classical Gram–Schmidt to obtain
V =
_
_
_
1.0000e + 00 1.0436e −07 9.9715e −08
1.0000e −07 1.0456e −14 −9.9905e −01
1.0000e −07 −1.0000e + 00 4.3568e −02
_
_
_.
The columns of V
U
are not orthogonal at all. In fact v
T
2
v
3
≈ −.004. For
modified Gram–Schmidt
V =
_
_
_
1.0000e + 00 1.0436e −07 1.0436e −07
1.0000e −07 1.0456e −14 −1.0000e + 00
1.0000e −07 −1.0000e + 00 4.3565e −16
_
_
_.
Here [v
T
i
v
j
−δ
ij
[ ≤ 10
−8
for all i, j.
The versions we implement in the collection of MATLAB codes use modi-
fied Gram–Schmidt. The outline of our implementation is Algorithm gmresb.
This implementation solves the upper Hessenberg least squares problem using
the MATLAB backward division operator, and is not particularly efficient. We
present a better implementation in Algorithm gmres. However, this version is
very simple and illustrates some important ideas. First, we see that x
k
need
only be computed after termination as the least squares residual ρ can be used
to approximate the norm of the residual (they are identical in exact arithmetic).
Second, there is an opportunity to compensate for a loss of orthogonality in
the basis vectors for the Krylov space. One can take a second pass through the
modified Gram–Schmidt process and restore lost orthogonality [147], [160].
Algorithm 3.4.3. gmresb(x, b, A, , kmax, ρ)
1. r = b −Ax, v
1
= r/|r|
2
, ρ = |r|
2
, β = ρ, k = 0
2. While ρ > |b|
2
and k < kmax do
(a) k = k + 1
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 41
(b) v
k+1
= Av
k
for j = 1, . . . k
i. h
jk
= v
T
k+1
v
j
ii. v
k+1
= v
k+1
−h
jk
v
j
(c) h
k+1,k
= |v
k+1
|
2
(d) v
k+1
= v
k+1
/|v
k+1
|
2
(e) e
1
= (1, 0, . . . , 0)
T
∈ R
k+1
Minimize |βe
1
−H
k
y
k
|
R
k+1 to obtain y
k
∈ R
k
.
(f) ρ = |βe
1
−H
k
y
k
|
R
k+1.
3. x
k
= x
0
+V
k
y
k
.
Even if modified Gram–Schmidt orthogonalization is used, one can still
lose orthogonality in the columns of V . One can test for loss of orthogonality
[22], [147], and reorthogonalize if needed or use a more stable means to create
the matrix V [195]. These more complex implementations are necessary if A is
ill conditioned or many iterations will be taken. For example, one can augment
the modified Gram–Schmidt process
• v
k+1
= Av
k
for j = 1, . . . k
h
jk
= v
T
k+1
v
j
v
k+1
= v
k+1
−h
jk
v
j
• h
k+1,k
= |v
k+1
|
2
• v
k+1
= v
k+1
/|v
k+1
|
2
with a second pass (reorthogonalization). One can reorthogonalize in every
iteration or only if a test [147] detects a loss of orthogonality. There is nothing
to be gained by reorthogonalizing more than once [147].
The modified Gram–Schmidt process with reorthogonalization looks like
• v
k+1
= Av
k
for j = 1, . . . , k
h
jk
= v
T
k+1
v
j
v
k+1
= v
k+1
−h
jk
v
j
• h
k+1,k
= |v
k+1
|
2
• If loss of orthogonality is detected
For j = 1, . . . , k
h
tmp
= v
T
k+1
v
j
h
jk
= h
jk
+h
tmp
v
k+1
= v
k+1
−h
tmp
v
j
• h
k+1,k
= |v
k+1
|
2
• v
k+1
= v
k+1
/|v
k+1
|
2
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
42 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
One approach to reorthogonalization is to reorthogonalize in every step.
This doubles the cost of the computation of V and is usually unnecessary.
More efficient and equally effective approaches are based on other ideas. A
variation on a method from [147] is used in [22]. Reorthogonalization is done
after the Gram–Schmidt loop and before v
k+1
is normalized if
|Av
k
|
2
+δ|v
k+1
|
2
= |Av
k
|
2
(3.10)
to working precision. The idea is that if the new vector is very small relative
to Av
k
then information may have been lost and a second pass through
the modified Gram–Schmidt process is needed. We employ this test in the
MATLAB code gmres with δ = 10
−3
.
To illustrate the effects of loss of orthogonality and those of reorthogonal-
ization we apply GMRES to the diagonal system Ax = b where b = (1, 1, 1)
T
,
x
0
= (0, 0, 0)
T
, and
A =
_
_
_
.001 0 0
0 .0011 0
0 0 10
4
_
_
_. (3.11)
While in infinite precision arithmetic only three iterations are needed to solve
the system exactly, we find in the MATLAB environment that a solution to full
precision requires more than three iterations unless reorthogonalization is ap-
plied after every iteration. In Table 3.1 we tabulate relative residuals as a func-
tion of the iteration counter for classical Gram–Schmidt without reorthogonal-
ization (CGM), modified Gram–Schmidt without reorthogonalization (MGM),
reorthogonalization based on the test (3.10) (MGM-PO), and reorthogonaliza-
tion in every iteration (MGM-FO). While classical Gram–Schmidt fails, the
reorthogonalization strategy based on (3.10) is almost as effective as the much
more expensive approach of reorthogonalizing in every step. The method based
on (3.10) is the default in the MATLAB code gmresa.
The kth GMRES iteration requires a matrix-vector product, k scalar
products, and the solution of the Hessenberg least squares problem in step 2e.
The k scalar products require O(kN) floating-point operations and the cost
of a solution of the Hessenberg least squares problem, by QR factorization or
the MATLAB backward division operator, say, in step 2e of gmresb is O(k
3
)
floating-point operations. Hence the total cost of the m GMRES iterations is
m matrix-vector products and O(m
4
+m
2
N) floating-point operations. When
k is not too large and the cost of matrix-vector products is high, a brute-
force solution to the least squares problem using the MATLAB backward
division operator is not terribly inefficient. We provide an implementation
of Algorithm gmresb in the collection of MATLAB codes. This is an appealing
algorithm, especially when implemented in an environment like MATLAB,
because of its simplicity. For large k, however, the brute-force method can be
very costly.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 43
Table 3.1
Effects of reorthogonalization.
k CGM MGM MGM-PO MGM-FO
0 1.00e+00 1.00e+00 1.00e+00 1.00e+00
1 8.16e-01 8.16e-01 8.16e-01 8.16e-01
2 3.88e-02 3.88e-02 3.88e-02 3.88e-02
3 6.69e-05 6.42e-08 6.42e-08 6.34e-34
4 4.74e-05 3.70e-08 5.04e-24
5 3.87e-05 3.04e-18
6 3.35e-05
7 3.00e-05
8 2.74e-05
9 2.53e-05
10 2.37e-05
3.5. Implementation: Givens rotations
If k is large, implementations using Givens rotations [167], [22], Householder
reflections [195], or a shifted Arnoldi process [197] are much more efficient
than the brute-force approach in Algorithm gmresb. The implementation in
Algorithm gmres and in the MATLAB code collection is from [167]. This
implementation maintains the QR factorization of H
k
in a clever way so that
the cost for a single GMRES iteration is O(Nk) floating-point operations. The
O(k
2
) cost of the triangular solve and the O(kN) cost of the construction of
x
k
are incurred after termination.
A 2 2 Givens rotation is a matrix of the form
G =
_
c −s
s c
_
, (3.12)
where c = cos(θ), s = sin(θ) for θ ∈ [−π, π]. The orthogonal matrix G rotates
the vector (c, −s), which makes an angle of −θ with the x-axis through an
angle θ so that it overlaps the x-axis.
G
_
c
−s
_
=
_
1
0
_
.
An N N Givens rotation replaces a 2 2 block on the diagonal of the
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
44 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
N N identity matrix with a 2 2 Givens rotation.
G =
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
1 0 . . . 0
0
.
.
.
.
.
.
.
.
. c −s
.
.
. s c 0
.
.
.
0 1
.
.
.
.
.
.
.
.
. 0
0 . . . 0 1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
. (3.13)
Our notation is that G
j
(c, s) is an N N givens rotation of the form (3.13)
with a 2 2 Givens rotation in rows and columns j and j + 1.
Givens rotations are used to annihilate single nonzero elements of matrices
in reduction to triangular form [89]. They are of particular value in reducing
Hessenberg matrices to triangular form and thereby solving Hessenberg least
squares problems such as the ones that arise in GMRES. This reduction can be
accomplished in O(N) floating-point operations and hence is far more efficient
than a solution by a singular value decomposition or a reduction based on
Householder transformations. This method is also used in the QR algorithm
for computing eigenvalues [89], [184].
Let H be an N M (N ≥ M) upper Hessenberg matrix with rank M.
We reduce H to triangular form by first multiplying the matrix by a Givens
rotation that annihilates h
21
(and, of course, changes h
11
and the subsequent
columns). We define G
1
= G
1
(c
1
, s
1
) by
c
1
= h
11
/
_
h
2
11
+h
2
21
and s
1
= −h
21
/
_
h
2
11
+h
2
21
. (3.14)
If we replace H by G
1
H, then the first column of H now has only a single
nonzero element h
11
. Similarly, we can now apply G
2
(c
2
, s
2
) to H, where
c
2
= h
22
/
_
h
2
22
+h
2
32
and s
2
= −h
32
/
_
h
2
22
+h
2
32
. (3.15)
and annihilate h
32
. Note that G
2
does not affect the first column of H.
Continuing in this way and setting
Q = G
N
. . . G
1
we see that QH = R is upper triangular.
A straightforward application of these ideas to Algorithm gmres would
solve the least squares problem by computing the product of k Givens rotations
Q, setting g = βQe
1
, and noting that
|βe
1
−H
k
y
k
|
R
k+1 = |Q(βe
1
−H
k
y
k
)|
R
k+1 = |g −R
k
y
k
|
R
k+1,
where R
k
is the k + 1 k triangular factor of the QR factorization of H
k
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 45
In the context of GMRES iteration, however, we can incrementally perform
the QR factorization of H as the GMRES iteration progresses [167]. To see
this, note that if R
k
= Q
k
H
k
and, after orthogonalization, we add the new
column h
k+2
to H
k
, we can update both Q
k
and R
k
by first multiplying h
k+2
by Q
k
(that is applying the first k Givens rotations to h
k+2
), then computing
the Givens rotation G
k+1
that annihilates the (k + 2)nd element of Q
k
h
k+2
,
and finally setting Q
k+1
= G
k+1
Q
k
and forming R
k+1
by augmenting R
k
with
G
k+1
Q
k
h
k+2
.
The MATLAB implementation of Algorithm gmres stores Q
k
by storing
the sequences ¦c
j
¦ and ¦s
j
¦ and then computing the action of Q
k
on a vector
x ∈ R
k+1
by applying G
j
(c
j
, s
j
) in turn to obtain
Q
k
x = G
k
(c
k
, s
k
) . . . G
1
(c
1
, s
1
)x.
We overwrite the upper triangular part of H
k
with the triangular part of the
QR factorization of H
k
in the MATLAB code. The MATLAB implementation
of Algorithm gmres uses (3.10) to test for loss of orthogonality.
Algorithm 3.5.1. gmres(x, b, A, , kmax, ρ)
1. r = b −Ax, v
1
= r/|r|
2
, ρ = |r|
2
, β = ρ,
k = 0; g = ρ(1, 0, . . . , 0)
T
∈ R
kmax+1
2. While ρ > |b|
2
and k < kmax do
(a) k = k + 1
(b) v
k+1
= Av
k
for j = 1, . . . k
i. h
jk
= v
T
k+1
v
j
ii. v
k+1
= v
k+1
−h
jk
v
j
(c) h
k+1,k
= |v
k+1
|
2
(d) Test for loss of orthogonality and reorthogonalize if necessary.
(e) v
k+1
= v
k+1
/|v
k+1
|
2
(f) i. If k > 1 apply Q
k−1
to the kth column of H.
ii. ν =
_
h
2
k,k
+h
2
k+1,k
.
iii. c
k
= h
k,k
/ν, s
k
= −h
k+1,k

h
k,k
= c
k
h
k,k
−s
k
h
k+1,k
, h
k+1,k
= 0
iv. g = G
k
(c
k
, s
k
)g.
(g) ρ = [(g)
k+1
[.
3. Set r
i,j
= h
i,j
for 1 ≤ i, j ≤ k.
Set (w)
i
= (g)
i
for 1 ≤ i ≤ k.
Solve the upper triangular system Ry
k
= w.
4. x
k
= x
0
+V
k
y
k
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
46 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
We close with an example of an implementation of GMRES(m) . This
implementation does not test for success and may, therefore, fail to terminate.
You are asked to repair this in exercise 7. Aside from the parameter m, the
arguments to Algorithm gmresm are the same as those for Algorithm gmres.
Algorithm 3.5.2. gmresm(x, b, A, , kmax, m, ρ)
1. gmres(x, b, A, , m, ρ)
2. k = m
3. While ρ > |b|
2
and k < kmax do
(a) gmres(x, b, A, , m, ρ)
(b) k = k +m
The storage costs of m iterations of gmres or of gmresm are the m + 2
vectors b, x, and ¦v
k
¦
m
k=1
.
3.6. Other methods for nonsymmetric systems
The methods for nonsymmetric linear systems that receive most attention in
this book, GMRES, CGNR, and CGNE, share the properties that they are easy
to implement, can be analyzed by a common residual polynomial approach, and
only terminate if an acceptable approximate solution has been found. CGNR
and CGNE have the disadvantages that a transpose-vector product must be
computed for each iteration and that the coefficient matrix A
T
A or AA
T
has
condition number the square of that of the original matrix. In ¸ 3.8 we give
an example of how this squaring of the condition number can lead to failure
of the iteration. GMRES needs only matrix-vector products and uses A alone,
but, as we have seen, a basis for /
k
must be stored to compute x
k
. Hence,
the storage requirements increase as the iteration progresses. For large and
ill-conditioned problems, it may be impossible to store enough basis vectors
and the iteration may have to be restarted. Restarting can seriously degrade
performance.
An ideal method would, like CG, only need matrix-vector products, be
based on some kind of minimization principle or conjugacy, have modest
storage requirements that do not depend on the number of iterations needed
for convergence, and converge in N iterations for all nonsingular A. However,
[74], methods based on short-term recurrences such as CG that also satisfy
minimization or conjugacy conditions cannot be constructed for general
matrices. The methods we describe in this section fall short of the ideal, but
can still be quite useful. We discuss only a small subset of these methods and
refer the reader to [12] and [78] for pointers to more of the literature on this
subject. All the methods we present in this section require two matrix-vector
products for each iteration.
Consistently with our implementation of GMRES, we take the view that
preconditioners will be applied externally to the iteration. However, as with
CG, these methods can also be implemented in a manner that incorporates
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 47
the preconditioner and uses the residual for the original system to control
termination.
3.6.1. Bi-CG. The earliest such method Bi-CG (Biconjugate gradient)
[122], [76], does not enforce a minimization principle; instead, the kth residual
must satisfy the bi-orthogonality condition
r
T
k
w = 0 for all w ∈ /
k
, (3.16)
where
/
k
= span(ˆ r
0
, A
T
ˆ r
0
, . . . , (A
T
)
k−1
ˆ r
0
)
is the Krylov space for A
T
and the vector ˆ r
0
. ˆ r
0
is a user-supplied vector and is
often set to r
0
. The algorithm gets its name because it produces sequences of
residuals ¦r
k
¦, ¦ˆ r
k
¦ and search directions ¦p
k
¦, ¦ˆ p
k
¦ such that bi-orthogonality
holds, i. e. ˆ r
T
k
r
l
= 0 if k ,= l and the search directions ¦p
k
¦ and ¦ˆ p
k
¦ satisfy
the bi-conjugacy property
ˆ p
T
k
Ap
l
= 0 if k ,= l.
In the symmetric positive definite case (3.16) is equivalent to the minimization
principle (2.2) for CG [89].
Using the notation of Chapter 2 and [191] we give an implementation of Bi-
CG making the choice ˆ r
0
= r
0
. This algorithmic description is explained and
motivated in more detail in [122], [76], [78], and [191]. We do not recommend
use of Bi-CG and present this algorithm only as a basis for discussion of some
of the properties of this class of algorithms.
Algorithm 3.6.1. bicg(x, b, A, , kmax)
1. r = b −Ax, ˆ r = r, ρ
0
= 1, ˆ p = p = 0, k = 0
2. Do While |r|
2
> |b|
2
and k < kmax
(a) k = k + 1
(b) ρ
k
= ˆ r
T
r, β = ρ
k

k−1
(c) p = r +βp, ˆ p = ˆ r +βˆ p
(d) v = Ap
(e) α = ρ
k
/(ˆ p
T
v)
(f) x = x +αp
(g) r = r −αv; ˆ r = ˆ r −αA
T
ˆ p
Note that if A = A
T
is spd, Bi-CG produces the same iterations as CG
(but computes everything except x twice). If, however, A is not spd, there is
no guarantee that ρ
k
in step 2b or ˆ p
T
Ap, the denominator in step 2e, will not
vanish. If either ρ
k−1
= 0 or ˆ p
T
Ap = 0 we say that a breakdown has taken
place. Even if these quantities are nonzero but very small, the algorithm can
become unstable or produce inaccurate results. While there are approaches
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
48 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
[80], [78], [148], [79], [81], to avoid some breakdowns in many variations of
Bi-CG, there are no methods that both limit storage and completely eliminate
the possibility of breakdown [74]. All of the methods we present in this section
can break down and should be implemented with that possibility in mind.
Once breakdown has occurred, one can restart the Bi-CG iteration or pass
the computation to another algorithm, such as GMRES. The possibility of
breakdown is small and certainly should not eliminate the algorithms discussed
below from consideration if there is not enough storage for GMRES to perform
well.
Breakdowns aside, there are other problems with Bi-CG. A transpose-
vector product is needed, which at the least will require additional program-
ming and may not be possible at all. The performance of the algorithm can
be erratic or even unstable with residuals increasing by several orders of mag-
nitude from one iteration to the next. Moreover, the effort in computing ˆ r
at each iteration is wasted in that ˆ r makes no contribution to x. However,
Bi-CG sometimes performs extremely well and the remaining algorithms in
this section represent attempts to capture this good performance and damp
the erratic behavior when it occurs.
We can compare the best-case performance of Bi-CG with that of GMRES.
Note that there is ¯ p
k
∈ T
k
such that both
r
k
= ¯ p
k
(A)r
0
and ˆ r
k
= ¯ p
k
(A
T
)ˆ r
0
. (3.17)
Hence, by the minimization property for GMRES
|r
GMRES
k
|
2
≤ |r
Bi−CG
k
|
2
,
reminding us that GMRES, if sufficient storage is available, will always
reduce the residual more rapidly than Bi-CG (in terms of iterations, but not
necessarily in terms of computational work). One should also keep in mind that
a single Bi-CG iteration requires two matrix-vector products and a GMRES
iterate only one, but that the cost of the GMRES iteration increases (in terms
of floating-point operations) as the iteration progresses.
3.6.2. CGS. A remedy for one of the problems with Bi-CG is the Conjugate
Gradient Squared (CGS) algorithm [180]. The algorithm takes advantage of
the fact that (3.17) implies that the scalar product ˆ r
T
r in Bi-CG (step 2b of
Algorithm bicg) can be represented without using A
T
as
r
T
k
ˆ r
k
= (¯ p
k
(A)r
0
)
T
(¯ p
k
(A
T
)ˆ r
0
) = (¯ p
k
(A)
2
r
0
)
T
ˆ r
0
.
The other references to A
T
can be eliminated in a similar fashion and an
iteration satisfying
r
k
= ¯ p
k
(A)
2
r
0
(3.18)
is produced, where ¯ p
k
is the same polynomial produced by Bi-CG and used in
(3.17). This explains the name, Conjugate Gradient Squared.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 49
The work used in Bi-CG to compute ˆ r is now used to update x. CGS
replaces the transpose-vector product with an additional matrix-vector product
and applies the square of the Bi-CG polynomial to r
0
to produce r
k
. This may,
of course, change the convergence properties for the worse and either improve
good convergence or magnify erratic behavior [134], [191]. CGS has the same
potential for breakdown as Bi-CG. Since ¯ p
2
k
∈ T
2k
we have
|r
GMRES
2k
|
2
≤ |r
CGS
k
|
2
. (3.19)
We present an algorithmic description of CGS that we will refer to in our
discussion of TFQMR in ¸ 3.6.4. In our description of CGS we will index the
vectors and scalars that would be overwritten in an implementation so that we
can describe some of the ideas in TFQMR later on. This description is taken
from [77].
Algorithm 3.6.2. cgs(x, b, A, , kmax)
1. x
0
= x; p
0
= u
0
= r
0
= b −Ax
2. v
0
= Ap
0
; ˆ r
0
= r
0
3. ρ
0
= ˆ r
T
0
r
0
; k = 0
4. Do While |r|
2
> |b|
2
and k < kmax
(a) k = k + 1
(b) σ
k−1
= ˆ r
T
0
v
k−1
(c) α
k−1
= ρ
k−1

k−1
(d) q
k
= u
k−1
−α
k−1
v
k−1
(e) x
k
= x
k−1

k−1
(u
k−1
+q
k
)
r
k
= r
k−1
−α
k−1
A(u
k−1
+q
k
)
(f) ρ
k
= ˆ r
T
0
r
k
; β
k
= ρ
k

k−1
u
k
= r
k

k
q
k
p
k
= u
k

k
(q
k

k
p
k−1
)
v
k
= Ap
k
Breakdowns take place when either ρ
k−1
or σ
k−1
vanish. If the algorithm
does not break down, then α
k−1
,= 0 for all k. One can show [180], [77], that
if ¯ p
k
is the residual polynomial from (3.17) and (3.18) then
¯ p
k
(z) = ¯ p
k−1
(z) −α
k−1
z¯ q
k−1
(z) (3.20)
where the auxiliary polynomials ¯ q
k
∈ T
k
are given by ¯ q
0
= 1 and for k ≥ 1 by
¯ q
k
(z) = ¯ p
k
(z) +β
k
¯ q
k−1
(z). (3.21)
We provide no MATLAB implementation of Bi-CG or CGS. Bi-CGSTAB
is the most effective representative of this class of algorithms.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
50 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
3.6.3. Bi-CGSTAB. The Bi-conjugate gradient stabilized method (Bi-
CGSTAB) [191] attempts to smooth the convergence of CGS by replacing
(3.18) with
r
k
= q
k
(A)p
k
(A)r
0
(3.22)
where
q
k
(z) =
k

i=1
(1 −ω
i
z).
The constant ω
i
is selected to minimize r
i
= q
i
(A)p
i
(A)r
0
as a function of ω
i
.
The examples in [191] indicate the effectiveness of this approach, which can
be thought of [12] as a blend of Bi-CG and GMRES(1). The performance in
cases where the spectrum has a significant imaginary part can be improved by
constructing q
k
to have complex conjugate pairs of roots, [98].
There is no convergence theory for Bi-CGSTAB and no estimate of the
residuals that is better than that for CGS, (3.19). In our description of the
implementation, which is described and motivated fully in [191], we use ˆ r
0
= r
0
and follow the notation of [191].
Algorithm 3.6.3. bicgstab(x, b, A, , kmax)
1. r = b −Ax, ˆ r
0
= ˆ r = r, ρ
0
= α = ω = 1, v = p = 0, k = 0, ρ
1
= ˆ r
T
0
r
2. Do While |r|
2
> |b|
2
and k < kmax
(a) k = k + 1
(b) β = (ρ
k

k−1
)(α/ω)
(c) p = r +β(p −ωv)
(d) v = Ap
(e) α = ρ
k
/(ˆ r
T
0
v)
(f) s = r −αv, t = As
(g) ω = t
T
s/|t|
2
2
, ρ
k+1
= −ωˆ r
T
0
t
(h) x = x +αp +ωs
(i) r = s −ωt
Note that the iteration can break down in steps 2b and 2e. We provide an
implementation of Algorithm bicgstab in the collection of MATLAB codes.
The cost in storage and in floating-point operations per iteration re-
mains bounded for the entire iteration. One must store seven vectors
(x, b, r, p, v, ˆ r
0
, t), letting s overwrite r when needed. A single iteration re-
quires four scalar products. In a situation where many GMRES iterations are
needed and matrix-vector product is fast, Bi-CGSTAB can have a much lower
average cost per iterate than GMRES. The reason for this is that the cost of
orthogonalization in GMRES can be much more than that of a matrix-vector
product if the dimension of the Krylov space is large. We will present such a
case in ¸ 3.8.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 51
3.6.4. TFQMR. Now we consider the quasi-minimal residual (QMR)
family of algorithms [80], [81], [77], [37], [202]. We will focus on the transpose-
free algorithm TFQMR proposed in [77], to illustrate the quasi-minimization
idea. All algorithms in this family minimize the norm of an easily computable
quantity q = f − Hz, a quasi-residual over z ∈ R
N
. The quasi-residual is
related to the true residual by a full-rank linear transformation r = Lq and
reduction of the residual can be approximately measured by reduction in q.
The specific formulation of q and L depends on the algorithm.
Returning to Algorithm cgs we define sequences
y
m
=
_
¸
_
¸
_
u
k−1
if m = 2k −1 is odd,
q
k
if m = 2k is even,
and
w
m
=
_
¸
_
¸
_
(¯ p
k
(A))
2
r
0
if m = 2k −1 is odd,
¯ p
k
(A)¯ p
k−1
(A)r
0
if m = 2k is even.
Here the sequences ¦¯ p
k
¦ and ¦¯ q
k
¦ are given by (3.20) and (3.21). Hence, the
CGS residual satisfies
r
CGS
k
= w
2k+1
.
We assume that CGS does not break down and, therefore, α
k
,= 0 for all
k ≥ 0. If we let ¸r| be the nearest integer less than or equal to a real r we
have
Ay
m
= (w
m
−w
m+1
)/α
(m−1)/2
, (3.23)
where the denominator is nonzero by assumption. We express (3.23) in matrix
form as
AY
m
= W
m+1
B
m
, (3.24)
where Y
m
is the N m matrix with columns ¦y
j
¦
m
j=1
, W
m+1
the N (m+1)
matrix with columns ¦w
j
¦
m+1
j=1
, and B
m
is the (m+ 1) m matrix given by
B
m
=
_
_
_
_
_
_
_
_
_
_
1 0 . . . 0
−1 1
.
.
.
.
.
.
0
.
.
.
.
.
. 0
.
.
.
.
.
. −1 1
0 . . . 0 −1
_
_
_
_
_
_
_
_
_
_
diag(α
0
, α
0
, . . . , α
(m−1)/2
)
−1
.
Our assumptions that no breakdown takes place imply that /
m
is the span
of ¦y
j
¦
m
j=1
and hence
x
m
= x
0
+Y
m
z (3.25)
for some z ∈ R
m
. Therefore,
r
m
= r
0
−AY
m
z = W
m+1
(e
1
−B
m
z), (3.26)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
52 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
where, as in our discussion of GMRES, e
1
= (1, 0, . . . , 0)
T
∈ R
m
. In a sense,
the conditioning of the matrix W
m+1
can be improved by scaling. Let
Ω = diag(ω
1
, ω
2
, . . . , ω
m
) (3.27)
and rewrite (3.26) as
r
m
= r
0
−AY
m
z = W
m+1

−1
m+1
(f
m+1
−H
m
z) (3.28)
where f
m+1
= Ω
m+1
e
1
= ω
m+1
e
1
and H
m
= Ω
m+1
B
m
is bidiagonal and hence
upper Hessenberg. The diagonal entries in the matrix Ω
m+1
are called weights.
The weights make it possible to derive straightforward estimates of the residual
in terms of easily computed terms in the iteration.
If W
m+1

−1
m+1
were an orthogonal matrix, we could minimize r
m
by solving
the least squares problem
minimize
z∈R
m|f
m+1
−H
m
z|
2
. (3.29)
The quasi-minimization idea is to solve (3.29) (thereby quasi-minimizing the
residual) to obtain z
m
and then set
x
m
= x
0
+Y
m
z
m
. (3.30)
Note that if k corresponds to the iteration counter for CGS, each k produces
two approximate solutions (corresponding to m = 2k −1 and m = 2k).
The solution to the least squares problem (3.29) can be done by using
Givens rotations to update the factorization of the upper Hessenberg matrix
H
m
[77] and this is reflected in the implementation description below. As
one can see from that description, approximate residuals are not computed in
Algorithm tfqmr. Instead we have the quasi-residual norm
τ
m
= |f
m+1
−H
m
z
m
|
2
.
Now if we pick the weights ω
i
= |w
i
|
2
then the columns of W
m+1

−1
m+1
have
norm one. Hence
|r
m
|
2
= |W
m+1

−1
m+1
(f
m+1
−H
m
z)|
2
≤ τ
m

m+ 1. (3.31)
We can, therefore, base our termination criterion on τ
m
, and we do this in
Algorithm tfqmr.
One can [80], [78], also use the quasi-minimization condition to estimate
the TFQMR residual in terms of residual polynomials.
Theorem 3.6.1. Let A be an N N matrix and let x
0
, b ∈ R
N
be given.
Assume that the TFQMR iteration does not break down or terminate with
the exact solution for k iterations and let 1 ≤ m ≤ 2k. Let r
GMRES
m
be the
residual for the mth GMRES iteration. Let ξ
m
be the smallest singular value
of W
m+1

−1
m+1
. Then if ξ
m
> 0
τ
m
≤ |r
GMRES
m
|
2

m
. (3.32)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 53
Proof. We may write
r
GMRES
m
= r
0
+Y
m
z
GMRES
m
and obtain
|r
GMRES
m
|
2
= |W
m+1

−1
m+1
(f
m+1
−H
m
z
GMRES
m
)|
2
≥ ξ
m
|f
m+1
−H
m
z
GMRES
m
|
2
.
Hence, by the quasi-minimization property
|r
GMRES
m
|
2
≥ ξ
m
τ
m
as desired.
As a corollary we have a finite termination result and a convergence
estimate for diagonalizable matrices similar to Theorem 3.1.3.
Corollary 3.6.1. Let A be an N N matrix and let x
0
, b ∈ R
N
be given.
Then within (N +1)/2 iterations the TFQMR iteration will either break down
or terminate with the solution.
We apply Theorem 3.1.3 to (3.32) and use (3.31) to obtain the following
result.
Theorem 3.6.2. Let A = V ΛV
−1
be a nonsingular diagonalizable matrix.
Assume that TFQMR does not break down or terminate with the solution for k
iterations. For 1 ≤ m ≤ 2k let ξ
m
be the smallest singular value of W
m+1

−1
m+1
and let x
m
be given by (3.30). Then, if ξ
m
> 0,
|r
m
|
2
|r
0
|
2


m+ 1ξ
−1
m
κ(V ) max
z∈σ(A)
[φ(z)[
for all φ ∈ T
m
.
The implementation below follows [77] with ˆ r
0
= r
0
used throughout.
Algorithm 3.6.4. tfqmr(x, b, A, , kmax)
1. k = 0, w
1
= y
1
= r
0
= b −Ax, u
1
= v = Ay
1
, d = 0
ρ
0
= r
T
0
r
0
, τ = |r|
2
, θ = 0, η = 0
2. Do While k < kmax
(a) k = k + 1
(b) σ
k−1
= r
T
0
v, α = ρ
k−1

k−1
, y
2
= y
1
−α
k−1
v, u
2
= Ay
2
(c) For j = 1, 2 (m = 2k −2 +j)
i. w = w −α
k−1
u
j
ii. d = y
j
+ (θ
2
η/α
k−1
)d
iii. θ = |w|
2
/τ, c = 1/

1 +θ
2
iv. τ = τθc, η = c
2
α
k−1
v. x = x +ηd
vi. If τ

m+ 1 ≤ |b|
2
terminate successfully
(d) ρ
k
= r
T
0
w, β = ρ
k

k−1
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
54 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
(e) y
1
= w +βy
2
, u
1
= Ay
1
(f) v = u
1
+β(u
2
+βv)
Note that y
2
and u
2
= Ay
2
need only be computed if the loop in step 2c
does not terminate when j = 1. We take advantage of this in the MATLAB
code tfqmr to potentially reduce the matrix-vector product cost by one.
3.7. Examples for GMRES iteration
These examples use the code gmres from the collection of MATLAB codes. The
inputs, described in the comment lines, are the initial iterate x
0
, the right-hand
side vector b, a MATLAB function for the matrix-vector product, and iteration
parameters to specify the maximum number of iterations and the termination
criterion. We do not pass the preconditioner to the GMRES code, rather
we pass the preconditioned problem. In all the GMRES examples, we limit
the number of GMRES iterations to 60. For methods like GMRES(m), Bi-
CGSTAB, CGNR, and TFQMR, whose storage requirements do not increase
with the number of iterations, we allow for more iterations.
We consider the discretization of the partial differential equation
(Lu)(x, y) = −(u
xx
(x, y) +u
yy
(x, y)) +a
1
(x, y)u
x
(x, y)
+a
2
(x, y)u
y
(x, y) +a
3
(x, y)u(x, y) = f(x, y)
(3.33)
on 0 < x, y < 1 subject to homogeneous Dirichlet boundary conditions
u(x, 0) = u(x, 1) = u(0, y) = u(1, y) = 0, 0 < x, y < 1.
For general coefficients ¦a
j
¦ the operator L is not self-adjoint and its discretiza-
tion is not symmetric. As in ¸ 2.7 we discretize with a five-point centered dif-
ference scheme with n
2
points and mesh width h = 1/(n + 1). The unknowns
are
u
ij
≈ u(x
i
, x
j
)
where x
i
= ih for 1 ≤ i ≤ n. We compute Lu in a matrix-free manner
as we did in ¸ 2.7. We let n = 31 to create a system with 961 unknowns.
As before, we expect second-order accuracy and terminate the iteration when
|r
k
|
2
≤ h
2
|b|
2
≈ 9.8 10
−4
|b|
2
.
As a preconditioner we use the fast Poisson solver fish2d described in ¸ 2.7.
The motivation for this choice is similar to that for the conjugate gradient
computations and has been partially explained theoretically in [33], [34], [121],
and [139]. If we let Gu denote the action of the Poisson solver on u, the
preconditioned system is GLu = Gf.
For the computations reported in this section we took
a
1
(x, y) = 1, a
2
(x, y) = 20y, and a
3
(x, y) = 1.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 55
u
0
= 0 was the initial iterate. As in ¸ 2.7 we took the right hand side so that
the exact solution was the discretization of
10xy(1 −x)(1 −y) exp(x
4.5
).
In Figs. 3.1 and 3.2 we plot iteration histories corresponding to precondi-
tioned or unpreconditioned GMRES. For the restarted iteration, we restarted
the algorithm after three iterations, GMRES(3) in the terminology of ¸ 3.4.
Note that the plots are semilog plots of |r|
2
/|b|
2
and that the right-hand
sides in the preconditioned and unpreconditioned cases are not the same. In
both figures the solid line is the history of the preconditioned iteration and the
dashed line that of the unpreconditioned. Note the importance of precondi-
tioning. The MATLAB flops command indicates that the unpreconditioned
iteration required 7.6 million floating point operations and converged in 56
iterations. The preconditioned iteration required 1.3 million floating point op-
erations and terminated after 8 iterations. Restarting after 3 iterations is more
costly in terms of the iteration count than not restarting. However, even in
the unpreconditioned case, the restarted iteration terminated successfully. In
this case the MATLAB flops command indicates that the unpreconditioned
iteration terminated after 223 iterations and 9.7 million floating point opera-
tions and the preconditioned iteration required 13 iterations and 2.6 million
floating point operations.
The execution times in our environment indicate that the preconditioned
iterations are even faster than the difference in operation counts indicates.
The reason for this is that the FFT routine is a MATLAB intrinsic and is not
interpreted MATLAB code. In other computing environments preconditioners
which vectorize or parallelize particularly well might perform in the same way.
3.8. Examples for CGNR, Bi-CGSTAB, and TFQMR iteration
When a transpose is available, CGNR (or CGNE) is an alternative to GMRES
that does not require storage of the iteration history. For elliptic problems like
(3.33), the transpose (adjoint) of L is given by
L

u = −u
xx
−u
yy
−a
1
u
x
−a
2
u
y
+a
3
u
as one can derive by integration by parts. To apply CGNR to the precondi-
tioned problem, we require the transpose of GL, which is given by
(GL)

u = L

G

u = L

Gu.
The cost of the application of the transpose of L or GL is the same as that
of the application of L or GL itself. Hence, in the case where the cost of the
matrix-vector product dominates the computation, a single CGNR iteration
costs roughly the same as two GMRES iterations.
However, CGNR has the effect of squaring the condition number, this effect
has serious consequences as Fig. 3.3 shows. The dashed line corresponds to
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
56 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
0 10 20 30 40 50 60
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

R
e
s
i
d
u
a
l
Fig. 3.1. GMRES for 2-D elliptic equation.
0 50 100 150 200 250
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

R
e
s
i
d
u
a
l
Fig. 3.2. GMRES(3) for 2-D elliptic equation.
CGNR on the unpreconditioned problem, i.e., CG applied to L

Lu = L

f.
The matrix corresponding to the discretization of L

L has a large condition
number. The solid line corresponds to the CGNR applied to the preconditioned
problem, i.e., CG applied to L

G
2
Lu = L

G
2
f. We limited the number of
iterations to 310 = 10m and the unpreconditioned formulation of CGNR had
made very little progress after 310 iterations. This is a result of squaring
the large condition number. The preconditioned formulation did quite well,
converging in eight iterations. The MATLAB flops command indicates that
the unpreconditioned iteration required 13.7 million floating-point operations
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 57
and the preconditioned iteration 2.4 million. However, as you are asked to
investigate in Exercise 3.9.8, the behavior of even the preconditioned iteration
can also be poor if the coefficients of the first derivative terms are too large.
CGNE has similar performance; see Exercise 3.9.9.
0 50 100 150 200 250 300 350
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

R
e
s
i
d
u
a
l
Fig. 3.3. CGNR for 2-D elliptic equation.
The examples for Bi-CGSTAB use the code bicgstab from the collection
of MATLAB codes. This code has the same input/output arguments as gmres.
We applied Bi-CGSTAB to both the preconditioned and unpreconditioned
forms of (3.33) with the same data and termination criterion as for the GMRES
example.
We plot the iteration histories in Fig. 3.4. As in the previous examples, the
solid line corresponds to the preconditioned iteration and the dashed line to
the unpreconditioned iteration. Neither convergence history is monotonic, with
the unpreconditioned iteration being very irregular, but not varying by many
orders of magnitude as a CGS or Bi-CG iteration might. The preconditioned
iteration terminated in six iterations (12 matrix-vector products) and required
roughly 1.7 million floating-point operations. The unpreconditioned iteration
took 40 iterations and 2.2 million floating-point operations. We see a
considerable improvement in cost over CGNR and, since our unpreconditioned
matrix-vector product is so inexpensive, a better cost/iterate than GMRES in
the unpreconditioned case.
We repeat the experiment with TFQMR as our linear solver. We use
the code tfqmr that is included in our collection. In Fig. 3.5 we plot the
convergence histories in terms of the approximate residual

2k + 1τ
2k
given
by (3.31). We plot only full iterations (m = 2k) except, possibly, for the
final iterate. The approximate residual is given in terms of the quasi-residual
norm τ
m
rather than the true residual, and the graphs are monotonic. Our
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
58 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
0 5 10 15 20 25 30 35 40
10
-4
10
-3
10
-2
10
-1
10
0
10
1
Iterations
R
e
l
a
t
i
v
e

R
e
s
i
d
u
a
l
Fig. 3.4. Bi-CGSTAB for 2-D elliptic equation.
0 10 20 30 40 50 60 70
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

A
p
p
r
o
x
i
m
a
t
e

R
e
s
i
d
u
a
l
Fig. 3.5. TFQMR for 2-D elliptic equation.
termination criterion is based on the quasi-residual and (3.31) as described in
¸ 3.6.4, stopping the iteration when τ
m

m+ 1/|b|
2
≤ h
−2
. This led to actual
relative residuals of 7 10
−5
for the unpreconditioned iteration and 2 10
−4
for the preconditioned iteration, indicating that it is reasonable to use the
quasi-residual estimate (3.31) to make termination decisions.
The unpreconditioned iteration required roughly 4.1 million floating-point
operations and 67 iterations for termination. The preconditioned iteration
was much more efficient, taking 1.9 million floating-point operations and 7
iterations. The plateaus in the graph of the convergence history for the
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 59
unpreconditioned iteration are related (see [196]) to spikes in the corresponding
graph for a CGS iteration.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
60 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
3.9. Exercises on GMRES
3.9.1. Give an example of a matrix that is not diagonalizable.
3.9.2. Prove Theorem 3.1.4 and Theorem 3.1.5.
3.9.3. Prove that the Arnoldi process (unless it terminates prematurely with a
solution) produces a basis for /
k+1
that satisfies (3.8).
3.9.4. Duplicate Table 3.1 and add two more columns corresponding to classical
Gram–Schmidt orthogonalization with reorthogonalization at every step
and determine when the test in (3.10) detects loss of orthogonality. How
do your conclusions compare with those in [160]?
3.9.5. Let A = J
λ
be an elementary Jordan block of size N corresponding to
an eigenvalue λ ,= 0. This means that
J
λ
=
_
_
_
_
_
_
_
_
_
_
λ 1 0
0 λ 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 λ 1 0
0 λ 1
0 λ
_
_
_
_
_
_
_
_
_
_
.
Give an example of x
0
, b ∈ R
N
, and λ ∈ R such that the GMRES
residuals satisfy |r
k
|
2
= |r
0
|
2
for all k < N.
3.9.6. Consider the centered difference discretization of
−u

+u

+u = 1, u(0) = u(1) = 0.
Solve this problem with GMRES (without preconditioning) and then
apply as a preconditioner the map M defined by
−(Mf)

= f, (Mf)(0) = (Mf)(1) = 0.
That is, precondition with a solver for the high-order term in the
differential equation using the correct boundary conditions. Try this
for meshes with 50, 100, 200, . . . points. How does the performance of
the iteration depend on the mesh? Try other preconditioners such as
Gauss–Seidel and Jacobi. How do other methods such as CGNR, CGNE,
Bi-CGSTAB, and TFQMR perform?
3.9.7. Modify the MATLAB GMRES code to allow for restarts. One way to do
this is to call the gmres code with another code which has an outer loop
to control the restarting as in Algorithm gmresm. See [167] for another
example. How would you test for failure in a restarted algorithm? What
kinds of failures can occur? Repeat Exercise 6 with a limit of three
GMRES iterations before a restart.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GMRES ITERATION 61
3.9.8. Duplicate the results in ¸ 3.7 and ¸ 3.8. Compare the execution times
in your computing environment. Experiment with different values of m.
Why might GMRES(m) be faster than GMRES? Vary the dimension and
the coefficient a
y
. Try a
y
= 5y. Why is the performance of the methods
sensitive to a
y
and a
x
?
3.9.9. Apply CGNE to the PDE in ¸ 3.7. How does the performance differ from
CGNR and GMRES?
3.9.10. Consider preconditioning from the right. For the PDE in ¸ 3.7, apply
GMRES, CGNR, Bi-CGSTAB, TFQMR, and/or CGNE, to the equation
LGw = f and then recover the solution from u = Gw. Compare timings
and solution accuracy with those obtained by left preconditioning.
3.9.11. Are the solutions to (3.33) in ¸ 3.7 and ¸ 3.8 equally accurate? Compare
the final results with the known solution.
3.9.12. Implement Bi-CG and CGS and try them on some of the examples in
this chapter.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
62 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 4
Basic Concepts and Fixed-Point Iteration
This part of the book is about numerical methods for the solution of systems
of nonlinear equations. Throughout this part, we will let | | denote a norm
on R
N
as well as the induced matrix norm.
We begin by setting the notation. We seek to solve
F(x) = 0. (4.1)
Here F : R
N
→ R
N
. We denote the ith component of F by f
i
. If the
components of F are differentiable at x ∈ R
N
we define the Jacobian matrix
F

(x) by
F

(x)
ij
=
∂f
i
∂x
j
(x).
The Jacobian matrix is the vector analog of the derivative. We may express
the fundamental theorem of calculus as follows.
Theorem 4.0.1. Let F be differentiable in an open set Ω ⊂ R
N
and let
x

∈ Ω. Then for all x ∈ Ω sufficiently near x

F(x) −F(x

) =
_
1
0
F

(x

+t(x −x

))(x −x

) dt.
4.1. Types of convergence
Iterative methods can be classified by the rate of convergence.
Definition 4.1.1. Let ¦x
n
¦ ⊂ R
N
and x

∈ R
N
. Then
• x
n
→ x

q-quadratically if x
n
→ x

and there is K > 0 such that
|x
n+1
−x

| ≤ K|x
n
−x

|
2
.
• x
n
→ x

q-superlinearly with q-order α > 1 if x
n
→ x

and there is
K > 0 such that
|x
n+1
−x

| ≤ K|x
n
−x

|
α
.
• x
n
→ x

q-superlinearly if
lim
n→∞
|x
n+1
−x

|
|x
n
−x

|
= 0.
65
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
66 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
• x
n
→ x

q-linearly with q-factor σ ∈ (0, 1) if
|x
n+1
−x

| ≤ σ|x
n
−x

|
for n sufficiently large.
Definition 4.1.2. An iterative method for computing x

is said to be
locally (q-quadratically, q-superlinearly, q-linearly, etc.) convergent if the
iterates converge to x

(q-quadratically, q-superlinearly, q-linearly, etc.) given
that the initial data for the iteration is sufficiently good.
Note that a q-superlinearly convergent sequence is also q-linearly conver-
gent with q-factor σ for any σ > 0. A q-quadratically convergent sequence is
q-superlinearly convergent with q-order 2.
In general, a q-superlinearly convergent method is preferable to a q-linearly
convergent one if the cost of a single iterate is the same for both methods. We
will see that often a method that is more slowly convergent, in terms of the
convergence types defined above, can have a cost/iterate so low that the slow
iteration is more efficient.
Sometimes errors are introduced into the iteration that are independent of
the errors in the approximate solution. An example of this would be a problem
in which the nonlinear function F is the output of another algorithm that
provides error control, such as adjustment of the mesh size in an approximation
of a differential equation. One might only compute F to low accuracy in the
initial phases of the iteration to compute a solution and tighten the accuracy
as the iteration progresses. The r-type convergence classification enables us to
describe that idea.
Definition 4.1.3. Let ¦x
n
¦ ⊂ R
N
and x

∈ R
N
. Then ¦x
n
¦ converges
to x

r-(quadratically, superlinearly, linearly) if there is a sequence ¦ξ
n
¦ ⊂ R
converging q-(quadratically, superlinearly, linearly) to zero such that
|x
n
−x

| ≤ ξ
n
.
We say that ¦x
n
¦ converges r-superlinearly with r-order α > 1 if ξ
n
→ 0
q-superlinearly with q-order α.
4.2. Fixed-point iteration
Many nonlinear equations are naturally formulated as fixed-point problems
x = K(x) (4.2)
where K, the fixed-point map, may be nonlinear. A solution x

of (4.2) is
called a fixed point of the map K. Such problems are nonlinear analogs of the
linear fixed-point problems considered in Chapter 1. In this section we analyze
convergence of fixed-point iteration, which is given by
x
n+1
= K(x
n
). (4.3)
This iterative method is also called nonlinear Richardson iteration, Picard
iteration, or the method of successive substitution.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
FIXED-POINT ITERATION 67
Before discussing convergence of fixed-point iteration we make two defini-
tions.
Definition 4.2.1. Let Ω ⊂ R
N
and let G : Ω → R
M
. G is Lipschitz
continuous on Ω with Lipschitz constant γ if
|G(x) −G(y)| ≤ γ|x −y|
for all x, y ∈ Ω.
Definition 4.2.2. Let Ω ⊂ R
N
. K : Ω → R
N
is a contraction mapping
on Ω if K is Lipschitz continuous on Ω with Lipschitz constant γ < 1.
The standard result for fixed-point iteration is the Contraction Mapping
Theorem [9]. Compare the proof to that of Lemma 1.2.1 and Corollary 1.2.1.
Theorem 4.2.1. Let Ω be a closed subset of R
N
and let K be a contraction
mapping on Ω with Lipschitz constant γ < 1 such that K(x) ∈ Ω for all x ∈ Ω.
Then there is a unique fixed point of K, x

∈ Ω, and the iteration defined by
(4.3) converges q-linearly to x

with q-factor γ for all initial iterates x
0
∈ Ω.
Proof. Let x
0
∈ Ω. Note that ¦x
n
¦ ⊂ Ω because x
0
∈ Ω and K(x) ∈ Ω
whenever x ∈ Ω. The sequence ¦x
n
¦ remains bounded since for all i ≥ 1
|x
i+1
−x
i
| = |K(x
i
) −K(x
i−1
)| ≤ γ|x
i
−x
i−1
| . . . ≤ γ
i
|x
1
−x
0
|,
and therefore
|x
n
−x
0
| = |

n−1
i=0
x
i+1
−x
i
|

n−1
i=0
|x
i+1
−x
i
| ≤ |x
1
−x
0
|

n−1
i=0
γ
i
≤ |x
1
−x
0
|/(1 −γ).
Now, for all n, k ≥ 0,
|x
n+k
−x
n
| = |K(x
n+k−1
) −K(x
n−1
)|
≤ γ|x
n+k−1
−x
n−1
|
≤ γ|K(x
n+k−2
) −K(x
n−2
)|
≤ γ
2
|x
n+k−2
−x
n−2
| ≤ . . . ≤ γ
n
|x
k
−x
0
|
≤ γ
n
|x
1
−x
0
|/(1 −γ).
Hence
lim
n,k→∞
|x
n+k
−x
n
| = 0
and therefore the sequence ¦x
n
¦ is a Cauchy sequence and has a limit x

[162].
If K has two fixed points x

and y

in Ω, then
|x

−y

| = |K(x

) −K(y

)| ≤ γ|x

−y

|.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
68 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Since γ < 1 this can only hold if |x

−y

| = 0, i. e. x

= y

. Hence the fixed
point is unique.
Finally we note that
|x
n+1
−x

| = |K(x
n
) −K(x

)| ≤ γ|x
n
−x

|,
which completes the proof.
One simple application of this theorem is to numerical integration of
ordinary differential equations by implicit methods. For example, consider
the initial value problem
y

= f(y), y(t
0
) = y
0
and its solution by the backward Euler method with time step h,
y
n+1
−y
n
= hf(y
n+1
). (4.4)
In (4.4) y
k
denotes the approximation of y(t
0
+ kh). Advancing in time from
y
n
to y
n+1
requires solution of the fixed-point problem
y = K(y) = y
n
+hf(y).
For simplicity we assume that f is bounded and Lipschitz continuous, say
[f(y)[ ≤ M and [f(x) −f(y)[ ≤ M[x −y[
for all x, y.
If hM < 1, we may apply the contraction mapping theorem with Ω = R
since for all x, y
|K(x) −K(y)| = h[f(x) −f(y)[ ≤ hM[x −y[.
From this we conclude that for h sufficiently small we can integrate forward
in time and that the time step h can be chosen independently of y. This time
step may well be too small to be practical, and the methods based on Newton’s
method which we present in the subsequent chapters are used much more often
[83]. This type of analysis was used by Picard to prove existence of solutions
of initial value problems [152]. Nonlinear variations of the classical stationary
iterative methods such as Gauss–Seidel have also been used; see [145] for a
more complete discussion of these algorithms.
4.3. The standard assumptions
We will make the standard assumptions on F.
Assumption 4.3.1.
1. Equation 4.1 has a solution x

.
2. F

: Ω → R
N×N
is Lipschitz continuous with Lipschitz constant γ.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
FIXED-POINT ITERATION 69
3. F

(x

) is nonsingular.
These assumptions can be weakened [108] without sacrificing convergence
of the methods we consider here. However the classical result on quadratic
convergence of Newton’s method requires them and we make them throughout.
Exercise 5.7.1 illustrates one way to weaken the standard assumptions.
Throughout this part, we will always denote a root of F by x

. We let
B(r) denote the ball of radius r about x

B(r) = ¦x[ |e| < r¦,
where
e = x −x

.
The notation introduced above will be used consistently. If x
n
is the nth iterate
of a sequence, e
n
= x
n
−x

is the error in that iterate.
Lemma 4.3.1 is an important consequence of the standard assumptions.
Lemma 4.3.1. Assume that the standard assumptions hold. Then there is
δ > 0 so that for all x ∈ B(δ)
|F

(x)| ≤ 2|F

(x

)|, (4.5)
|F

(x)
−1
| ≤ 2|F

(x

)
−1
|, (4.6)
and
|F

(x

)
−1
|
−1
|e|/2 ≤ |F(x)| ≤ 2|F

(x

)||e|. (4.7)
Proof. Let δ be small enough so that B(δ) ⊂ Ω. For all x ∈ B(δ) we have
|F

(x)| ≤ |F

(x

)| +γ|e|.
Hence (4.5) holds if γδ < |F

(x

)|.
The next result (4.6) is a direct consequence of the Banach Lemma if
|I −F

(x

)
−1
F

(x)| < 1/2.
This will follow from
δ <
|F

(x

)
−1
|
−1

since then
|I −F

(x

)
−1
F

(x)| = |F

(x

)
−1
(F

(x

) −F

(x))|
≤ γ|F

(x

)
−1
||e| ≤ γδ|F

(x

)
−1
| < 1/2.
(4.8)
To prove the final inequality (4.7), we note that if x ∈ B(δ) then x

+te ∈
B(δ) for all 0 ≤ t ≤ 1. We use (4.5) and Theorem 4.0.1 to obtain
|F(x)| ≤
_
1
0
|F

(x

+te)||e| dt ≤ 2|F

(x

)||e|
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
70 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
which is the right inequality in (4.7).
To prove the left inequality in (4.7) note that
F

(x

)
−1
F(x) = F

(x

)
−1
_
1
0
F

(x

+te)e dt
= e −
_
1
0
(I −F

(x

)
−1
F

(x

+te))e dt,
and hence, by (4.8)
|F

(x

)
−1
F(x)| ≥ |e|(1 −|
_
1
0
I −F

(x

)
−1
F

(x

+te)dt|) ≥ |e|/2.
Therefore
|e|/2 ≤ |F

(x

)
−1
F(x)| ≤ |F

(x

)
−1
||F(x)|,
which completes the proof.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 5
Newton’s Method
5.1. Local convergence of Newton’s method
We will describe iterative methods for nonlinear equations in terms of the
transition from a current iterate x
c
to a new iterate x
+
. In this language,
Newton’s method is
x
+
= x
c
−F

(x
c
)
−1
F(x
c
). (5.1)
We may also view x
+
as the root of the two-term Taylor expansion or linear
model of F about x
c
M
c
(x) = F(x
c
) +F

(x
c
)(x −x
c
).
In the context of single systems this method appeared in the 17th century
[140], [156], [13], [137].
The convergence result on Newton’s method follows from Lemma 4.3.1.
Theorem 5.1.1. Let the standard assumptions hold. Then there are K > 0
and δ > 0 such that if x
c
∈ B(δ) the Newton iterate from x
c
given by (5.1)
satisfies
|e
+
| ≤ K|e
c
|
2
. (5.2)
Proof. Let δ be small enough so that the conclusions of Lemma 4.3.1 hold.
By Theorem 4.0.1
e
+
= e
c
−F

(x
c
)
−1
F(x
c
) = F

(x
c
)
−1
_
1
0
(F

(x
c
) −F

(x

+te
c
))e
c
dt.
By Lemma 4.3.1 and the Lipschitz continuity of F

|e
+
| ≤ (2|F

(x

)
−1
|)γ|e
c
|
2
/2.
This completes the proof of (5.2) with K = γ|F

(x

)
−1
|.
The proof of convergence of the complete Newton iteration will be complete
if we reduce δ if needed so that Kδ < 1.
Theorem 5.1.2. Let the standard assumptions hold. Then there is δ such
that if x
0
∈ B(δ) the Newton iteration
x
n+1
= x
n
−F

(x
n
)
−1
F(x
n
)
converges q-quadratically to x

.
71
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
72 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Proof. Let δ be small enough so that the conclusions of Theorem 5.1.1
hold. Reduce δ if needed so that Kδ = η < 1. Then if n ≥ 0 and x
n
∈ B(δ),
Theorem 5.1.1 implies that
|e
n+1
| ≤ K|e
n
|
2
≤ η|e
n
| < |e
n
| (5.3)
and hence x
n+1
∈ B(ηδ) ⊂ B(δ). Therefore if x
n
∈ B(δ) we may continue the
iteration. Since x
0
∈ B(δ) by assumption, the entire sequence ¦x
n
¦ ⊂ B(δ).
(5.3) then implies that x
n
→ x

q-quadratically.
The assumption that the initial iterate be “sufficiently near” the solution
(x
0
∈ B(δ)) may seem artificial at first look. There are, however, many
situations in which the initial iterate is very near the root. Two examples are
implicit integration of ordinary differential equations and differential algebraic
equations, [105], [83], [16], where the initial iterate is derived from the solution
at the previous time step, and solution of discretizations of partial differential
equations, where the initial iterate is an interpolation of a solution from a
coarser computational mesh [99], [126]. Moreover, when the initial iterate is
far from the root and methods such as those discussed in Chapter 8 are needed,
local convergence results like those in this and the following two chapters
describe the terminal phase of the iteration.
5.2. Termination of the iteration
We can base our termination decision on Lemma 4.3.1. If x ∈ B(δ), where δ
is small enough so that the conclusions of Lemma 4.3.1 hold, then if F

(x

) is
well conditioned, we may terminate the iteration when the relative nonlinear
residual |F(x)|/|F(x
0
)| is small. We have, as a direct consequence of applying
Lemma 4.3.1 twice, a nonlinear form of Lemma 1.1.1.
Lemma 5.2.1. Assume that the standard assumptions hold. Let δ > 0 be
small enough so that the conclusions of Lemma 4.3.1 hold for x ∈ B(δ). Then
for all x ∈ B(δ)
|e|
4|e
0
|κ(F

(x

))

|F(x)|
|F(x
0
)|

4κ(F

(x

))|e|
|e
0
|
,
where κ(F

(x

)) = |F

(x

)||F

(x

)
−1
| is the condition number of F

(x

)
relative to the norm | |.
From Lemma 5.2.1 we conclude that if F

(x

) is well conditioned, the size
of the relative nonlinear residual is a good indicator of the size of the error.
However, if there is error in the evaluation of F or the initial iterate is near a
root, a termination decision based on the relative residual may be made too
late in the iteration or, as we will see in the next section, the iteration may
not terminate at all. We raised this issue in the context of linear equations
in Chapter 1 when we compared (1.2) and (1.4). In the nonlinear case, there
is no “right-hand side” to use as a scaling factor and we must balance the
relative and absolute size of the nonlinear residuals in some other way. In all
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 73
of our MATLAB codes and algorithms for nonlinear equations our termination
criterion is to stop the iteration if
|F(x)| ≤ τ
r
|F(x
0
)| +τ
a
, (5.4)
where the relative error tolerance τ
r
and absolute error tolerance τ
a
are input
to the algorithm. Combinations of relative and absolute error tolerances are
commonly used in numerical methods for ordinary differential equations or
differential algebraic equations [16], [21], [23], [151].
Another way to decide whether to terminate is to look at the Newton step
s = −F

(x
c
)
−1
F(x
c
) = x
+
−x
c
,
and terminate the iteration when |s| is sufficiently small. This criterion is
based on Theorem 5.1.1, which implies that
|e
c
| = |s| +O(|e
c
|
2
).
Hence, near the solution s and e
c
are essentially the same size.
For methods other than Newton’s method, the relation between the step
and the error is less clear. If the iteration is q-linearly convergent, say, then
|e
+
| ≤ σ|e
c
|
implies that
(1 −σ)|e
c
| ≤ |s| ≤ (1 +σ)|e
c
|.
Hence the step is a reliable indicator of the error provided σ is not too near
1. The differential equations codes discussed in [16], [21], [23], and [151], make
use of this in the special case of the chord method.
However, in order to estimate |e
c
| this way, one must do most of the work
needed to compute x
+
, whereas by terminating on small relative residuals or
on a condition like (5.4) the decision to stop the iteration can be made before
computation and factorization of F

(x
c
). If computation and factorization of
the Jacobian are inexpensive, then termination on small steps becomes more
attractive.
5.3. Implementation of Newton’s method
In this section we assume that direct methods will be used to solve the linear
equation for the Newton step. Our examples and discussions of operation
counts make the implicit assumption that the Jacobian is dense. However,
the issues for sparse Jacobians when direct methods are used to solve linear
systems are very similar. In Chapter 6 we consider algorithms in which iterative
methods are used to solve the equation for the Newton step.
In order to compute the Newton iterate x
+
from a current point x
c
one
must first evaluate F(x
c
) and decide whether to terminate the iteration. If
one decides to continue, the Jacobian F

(x
c
) must be computed and factored.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
74 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Then the step is computed as the solution of F

(x
c
)s = −F(x
c
) and the iterate
is updated x
+
= x
c
+s. Of these steps, the evaluation and factorization of F

are the most costly. Factorization of F

in the dense case costs O(N
3
) floating-
point operations. Evaluation of F

by finite differences should be expected to
cost N times the cost of an evaluation of F because each column in F

requires
an evaluation of F to form the difference approximation. Hence the cost of a
Newton step may be roughly estimated as N + 1 evaluations of F and O(N
3
)
floating-point operations. In many cases F

can be computed more efficiently,
accurately, and directly than with differences and the analysis above for the
cost of a Newton iterate is very pessimistic. See Exercise 5.7.21 for an example.
For general nonlinearities or general purpose codes such as the MATLAB code
nsol from the collection, there may be little alternative to difference Jacobians.
In the future, automatic differentiation (see [94] for a collection of articles on
this topic) may provide such an alternative.
For definiteness in the description of Algorithm newton we use an LU
factorization of F

. Any other appropriate factorization such as QR or
Cholesky could be used as well. The inputs of the algorithm are the initial
iterate x, the nonlinear map F, and a vector of termination tolerances
τ = (τ
r
, τ
a
) ∈ R
2
.
Algorithm 5.3.1. newton(x, F, τ)
1. r
0
= |F(x)|
2. Do while |F(x)| > τ
r
r
0

a
(a) Compute F

(x)
(b) Factor F

(x) = LU.
(c) Solve LUs = −F(x)
(d) x = x +s
(e) Evaluate F(x).
One approach to reducing the cost of items 2a and item 2b in the Newton
iteration is to move them outside of the main loop. This means that the linear
approximation of F(x) = 0 that is solved at each iteration has derivative
determined by the initial iterate.
x
+
= x
c
−F

(x
0
)
−1
F(x
c
).
This method is called the chord method. The inputs are the same as for
Newton’s method.
Algorithm 5.3.2. chord(x, F, τ)
1. r
0
= |F(x)|
2. Compute F

(x)
3. Factor F

(x) = LU.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 75
4. Do while |F(x)| > τ
r
r
0

a
(a) Solve LUs = −F(x)
(b) x = x +s
(c) Evaluate F(x).
The only difference in implementation from Newton’s method is that the
computation and factorization of the Jacobian are done before the iteration
is begun. The difference in the iteration itself is that an approximation to
F

(x
c
) is used. Similar differences arise if F

(x
c
) is numerically approximated
by differences. We continue in the next section with an analysis of the effects
of errors in the function and Jacobian in a general context and then consider
the chord method and difference approximation to F

as applications.
5.4. Errors in the function and derivative
Suppose that F and F

are computed inaccurately so that F + and F

+ ∆
are used instead of F and F

in the iteration. If ∆ is sufficiently small, the
resulting iteration can return a result that is an O() accurate approximation
to x

. This is different from convergence and was called “local improvement”
in [65]. These issues are discussed in [201] as well. If, for example, is entirely
due to floating-point roundoff, there is no reason to expect that |F(x
n
)| will
ever be smaller than in general. We will refer to this phase of the iteration
in which the nonlinear residual is no longer being reduced as stagnation of the
iteration.
Theorem 5.4.1. Let the standard assumptions hold. Then there are
¯
K > 0, δ > 0, and δ
1
> 0 such that if x
c
∈ B(δ) and |∆(x
c
)| < δ
1
then
x
+
= x
c
−(F

(x
c
) + ∆(x
c
))
−1
(F(x
c
) +(x
c
))
is defined (i.e., F

(x
c
) + ∆(x
c
) is nonsingular) and satisfies
|e
+
| ≤
¯
K(|e
c
|
2
+|∆(x
c
)||e
c
| +|(x
c
)|). (5.5)
Proof. Let δ be small enough so that the conclusions of Lemma 4.3.1 and
Theorem 5.1.1 hold. Let
x
N
+
= x
c
−F

(x
c
)
−1
F(x
c
)
and note that
x
+
= x
N
+
+ (F

(x
c
)
−1
−(F

(x
c
) + ∆(x
c
))
−1
)F(x
c
) −(F

(x
c
) + ∆(x
c
))
−1
(x
c
).
Lemma 4.3.1 and Theorem 5.1.1 imply
|e
+
| ≤ K|e
c
|
2
+ 2|F

(x
c
)
−1
−(F

(x
c
) + ∆(x
c
))
−1
)||F

(x

)||e
c
|
+|(F

(x
c
) + ∆(x
c
))
−1
||(x
c
)|.
(5.6)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
76 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
If
|∆(x
c
)| ≤ |F

(x

)
−1
|
−1
/4
then Lemma 4.3.1 implies that
|∆(x
c
)| ≤ |F

(x
c
)
−1
|
−1
/2
and the Banach Lemma implies that F

(x
c
) + ∆(x
c
) is nonsingular and
|(F

(x
c
) + ∆(x
c
))
−1
| ≤ 2|F

(x
c
)
−1
| ≤ 4|F

(x

)
−1
|.
Hence
|F

(x
c
)
−1
−(F

(x
c
) + ∆(x
c
))
−1
| ≤ 8|F

(x

)
−1
|
2
|∆(x
c
)|.
We use these estimates and (5.6) to obtain
|e
+
| ≤ K|e
c
|
2
+ 16|F

(x

)
−1
|
2
|F

(x

)||∆(x
c
)||e
c
| + 4|F

(x

)
−1
||(x
c
)|.
Setting
¯
K = K + 16|F

(x

)
−1
|
2
|F

(x

)| + 4|F

(x

)
−1
|
completes the proof.
The remainder of this section is devoted to applications of Theorem 5.4.1.
5.4.1. The chord method. Recall that the chord method is given by
x
+
= x
c
−F

(x
0
)
−1
F(x
c
).
In the language of Theorem 5.4.1
(x
c
) = 0, ∆(x
c
) = F

(x
0
) −F

(x
c
).
Hence, if x
c
, x
0
∈ B(δ) ⊂ Ω
|∆(x
c
)| ≤ γ|x
0
−x
c
| ≤ γ(|e
0
| +|e
c
|). (5.7)
We apply Theorem 5.4.1 to obtain the following result.
Theorem 5.4.2. Let the standard assumptions hold. Then there are
K
C
> 0 and δ > 0 such that if x
0
∈ B(δ) the chord iterates converge q-linearly
to x

and
|e
n+1
| ≤ K
C
|e
0
||e
n
|. (5.8)
Proof. Let δ be small enough so that B(δ) ⊂ Ω and the conclusions of
Theorem 5.4.1 hold. Assume that x
n
∈ B(δ). Combining (5.7) and (5.5)
implies
|e
n+1
| ≤
¯
K(|e
n
|(1 +γ) +γ|e
0
|)|e
n
| ≤
¯
K(1 + 2γ)δ|e
n
|.
Hence if δ is small enough so that
¯
K(1 + 2γ)δ = η < 1
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 77
then the chord iterates converge q-linearly to x

. Q-linear convergence implies
that |e
n
| < |e
0
| and hence (5.8) holds with K
C
=
¯
K(1 + 2γ).
Another variation of the chord method is
x
+
= x
c
−A
−1
F(x
c
),
where A ≈ F

(x

). Methods of this type may be viewed as preconditioned
nonlinear Richardson iteration. Since
|∆(x
c
)| = |A−F

(x
c
)| ≤ |A−F

(x

)| +|F

(x

) −F

(x
c
)|,
if x
c
∈ B(δ) ⊂ Ω then
|∆(x
c
)| ≤ |A−F

(x

)| +γ|e
c
| ≤ |A−F

(x

)| +γδ.
Theorem 5.4.3. Let the standard assumptions hold. Then there are
K
A
> 0, δ > 0, and δ
1
> 0, such that if x
0
∈ B(δ) and |A−F

(x

)| < δ
1
then
the iteration
x
n+1
= x
n
−A
−1
F(x
n
)
converges q-linearly to x

and
|e
n+1
| ≤ K
A
(|e
0
| +|A−F

(x

)|)|e
n
|. (5.9)
5.4.2. Approximate inversion of F

. Another way to implement chord-
type methods is to provide an approximate inverse of F

. Here we replace
F

(x)
−1
by B(x), where the action of B on a vector is less expensive to compute
than a solution using the LU factorization. Rather than express the iteration
in terms of B
−1
(x) −F

(x) and using Theorem 5.4.3 one can proceed directly
from the definition of approximate inverse.
Note that if B is constant (independent of x) then the iteration
x
+
= x
c
−BF(x
c
)
can be viewed as a preconditioned nonlinear Richardson iteration.
We have the following result.
Theorem 5.4.4. Let the standard assumptions hold. Then there are
K
B
> 0, δ > 0, and δ
1
> 0, such that if x
0
∈ B(δ) and the matrix-valued
function B(x) satisfies
|I −B(x)F

(x

)| = ρ(x) < δ
1
(5.10)
for all x ∈ B(δ) then the iteration
x
n+1
= x
n
−B(x
n
)F(x
n
)
converges q-linearly to x

and
|e
n+1
| ≤ K
B
(ρ(x
n
) +|e
n
|)|e
n
|. (5.11)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
78 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Proof. On one hand, this theorem could be regarded as a corollary of the
Banach Lemma and Theorem 5.4.1. We give a direct proof.
First, by (5.10) we have
|B(x)| = |B(x)F

(x

)F

(x

)
−1
| ≤ |B(x)F

(x

)||F

(x

)
−1
|
≤ M
B
= (1 +δ
1
)|F

(x

)
−1
|.
(5.12)
Using (5.12) and
e
+
= e
c
−B(x
c
)F(x
c
) =
_
1
0
(I −B(x
c
)F

(x

+te
c
))e
c
dt
= (I −B(x
c
)F

(x

))e
c
+B(x
c
)
_
1
0
(F

(x

) −F

(x

+te
c
))e
c
dt
we have
|e
+
| ≤ ρ(x
c
)|e
c
| +M
B
γ|e
c
|
2
/2.
This completes the proof with K
B
= 1 +M
B
γ/2.
5.4.3. The Shamanskii method. Alternation of a Newton step with a
sequence of chord steps leads to a class of high-order methods, that is, methods
that converge q-superlinearly with q-order larger than 2. We follow [18] and
name this method for Shamanskii [174], who considered the finite difference
Jacobian formulation of the method from the point of view of efficiency. The
method itself was analyzed in [190]. Other high-order methods, especially for
problems in one dimension, are described in [190] and [146].
We can describe the transition from x
c
to x
+
by
y
1
= x
c
−F

(x
c
)
−1
F(x
c
),
y
j+1
= y
j
−F

(x
c
)
−1
F(y
j
) for 1 ≤ j ≤ m−1,
x
+
= y
m
.
(5.13)
Note that m = 1 is Newton’s method and m = ∞ is the chord method with
¦y
j
¦ playing the role of the chord iterates.
Algorithmically a second loop surrounds the factor/solve step. The inputs
for the Shamanskii method are the same as for Newton’s method or the chord
method except for the addition of the parameter m. Note that we can overwrite
x
c
with the sequence of y
j
’s. Note that we apply the termination test after
computation of each y
j
.
Algorithm 5.4.1. sham(x, F, τ, m)
1. r
0
= |F(x)|
2. Do while |F(x)| > τ
r
r
0

a
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 79
(a) Compute F

(x)
(b) Factor F

(x) = LU.
(c) for j = 1, . . . m
i. Solve LUs = −F(x)
ii. x = x +s
iii. Evaluate F(x).
iv. If |F(x)| ≤ τ
r
r
0

a
exit.
The convergence result is a simple application of Theorem 5.4.2.
Theorem 5.4.5. Let the standard assumptions hold and let m ≥ 1 be
given. Then there are K
S
> 0 and δ > 0 such that if x
0
∈ B(δ) the Shamanskii
iterates converge q-superlinearly to x

with q-order m+ 1 and
|e
n+1
| ≤ K
S
|e
n
|
m+1
. (5.14)
Proof. Let δ be small enough so that B(δ) ⊂ Ω and that the conclusions
of Theorem 5.4.2 hold. Then if x
n
∈ B(δ) all the intermediate iterates ¦y
j
¦
are in B(δ) by Theorem 5.4.2. In fact, if we set y
1
= x
n
, (5.8) implies that for
1 ≤ j ≤ m
|y
j
−x

| ≤ K
C
|x
n
−x

||y
j−1
−x

| = K
C
|e
n
||y
j−1
−x

| ≤ . . . ≤ K
j
C
|e
n
|
j+1
.
Hence x
n+1
∈ B(δ). Setting j = m in the above inequality completes the
proof.
The advantage of the Shamanskii method over Newton’s method is
that high q-orders can be obtained with far fewer Jacobian evaluations or
factorizations. Optimal choices of m as a function of N can be derived
under assumptions consistent with dense matrix algebra [18] by balancing the
O(N
3
) cost of a matrix factorization with the cost of function and Jacobian
evaluation. The analysis in [18] and the expectation that, if the initial iterate is
sufficiently accurate, only a small number of iterations will be taken, indicate
that the chord method is usually the best option for very large problems.
Algorithm nsol is based on this idea and using an idea from [114] computes
and factors the Jacobian only until an estimate of the q-factor for the linear
convergence rate is sufficiently low.
5.4.4. Difference approximation to F

. Assume that we compute
F(x) + (x) instead of F(x) and attempt to approximate the action of F

on a vector by a forward difference. We would do this, for example, when
building an approximate Jacobian. Here the jth column of F

(x) is F

(x)e
j
where e
j
is the unit vector with jth component 1 and other components 0.
A forward difference approximation to F

(x)w would be
F(x +hw) +(x +hw) −F(x) −(x)
h
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
80 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Assume that |(x)| ≤ ¯ . Then
F

(x)w −
F(x +hw) +(x +hw) −F(x) −(x)
h
= O(h + ¯ /h).
The quantity inside the O-term is minimized when h =

¯ . This means, for
instance, that very small difference steps h can lead to inaccurate results. In
the special case where (x) is a result of floating-point roundoff in full precision
(¯ ≈ 10
−15
), the analysis here indicates that h ≈ 10
−7
is a reasonable choice.
The choice h ≈ 10
−15
in this case can lead to disaster (try it!). Here we have
made the implicit assumption that x and w are of roughly the same size. If
this is not the case, h should be scaled to reflect that. The choice
h = ¯
1/2
|x|/|w|
reflects all the discussion above.
A more important consequence of this analysis is that the choice of step
size in a forward difference approximation to the action of the Jacobian on
a vector must take into account the error in the evaluation of F. This can
become very important if part of F is computed from measured data.
If F(x) has already been computed, the cost of a forward difference
approximation of F

(x)w is an additional evaluation of F (at the point x+hw).
Hence the evaluation of a full finite difference Jacobian would cost N function
evaluations, one for each column of the Jacobian. Exploitation of special
structure could reduce this cost.
The forward difference approximation to the directional derivative is not
a linear operator in w. The reason for this is that the derivative has been
approximated, and so linearity has been lost. Because of this we must
carefully specify what we mean by a difference approximation to the Jacobian.
Definition 5.4.1 below specifies the usual choice.
Definition 5.4.1. Let F be defined in a neighborhood of x ∈ R
N
.
(∇
h
F)(x) is the N N matrix whose jth column is given by
(∇
h
F)(x)
j
=
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
F(x +h|x|e
j
) −F(x)
h|x|
x ,= 0
F(he
j
) −F(x)
h
x = 0
We make a similar definition of the difference approximation of the
directional derivative.
Definition 5.4.2. Let F be defined in a neighborhood of x ∈ R
N
and let
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 81
w ∈ R
N
. We have
D
h
F(x : w) =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
0, w = 0,
|w|
F(x +h|x|w/|w|) −F(x)
h|x|
, w, x ,= 0,
|w|
F(hw/|w|) −F(x)
h
, x = 0, w ,= 0.
(5.15)
The standard assumptions and the Banach Lemma imply the following
result.
Lemma 5.4.1. Let the standard assumptions hold. Then there are δ, ¯ ,
¯
h >
0 such that if x ∈ B(δ), h ≤
¯
h, and |(x)| ≤ ¯ for all x ∈ B(δ), then

h
(F +)(x) is nonsingular for all x ∈ B(δ) and there is M
F
> 0 such that
|F

(x) −∇
h
(F +)(x)| ≤ M
F
(h + ¯ /h).
If we assume that the forward difference step has been selected so that
h = O(

¯ ), then ∆(x) = O(

¯ ) by Lemma 5.4.1. Theorem 5.4.1 implies the
following result.
Theorem 5.4.6. Let the standard assumptions hold. Then there are
positive δ, ¯ , and K
D
such that if x
c
∈ B(δ), |(x)| ≤ ¯ for all x ∈ B(δ),
and there is M

> 0 such that
h
c
> M


¯
then
x
+
= x
c
−∇
hc
(F +)(x
c
)
−1
(F(x
c
) +(x
c
))
satisfies
|e
+
| ≤ K
D
(¯ + (|e
c
| +h
c
)|e
c
|).
Note that Theorem 5.4.6 does not imply that an iteration will converge to
x

. Even if is entirely due to floating-point roundoff, there is no guarantee
that (x
n
) → 0 as x
n
→ x

. Hence, one should expect the sequence ¦|e
n

to stagnate and cease to decrease once |e
n
| ≈ ¯ . Therefore in the early
phases of the iteration, while |e
n
| >>

¯ , the progress of the iteration will
be indistinguishable from Newton’s method as implemented with an exact
Jacobian. When |e
n
| ≤

¯ , Theorem 5.4.6 says that |e
n+1
| = O(¯ ). Hence
one will see quadratic convergence until the error in F admits no further
reduction.
A difference approximation to a full Jacobian does not use any special
structure of the Jacobian and will probably be less efficient than a hand coded
analytic Jacobian. If information on the sparsity pattern of the Jacobian is
available it is sometimes possible [47], [42], to evaluate ∇
h
F with far fewer than
N evaluations of F. If the sparsity pattern is such that F

can be reduced to
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
82 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
a block triangular form, the entire solution process can take advantage of this
structure [59].
For problems in which the Lipschitz constant of F

is large, the error in
the difference approximation will be large, too. Moreover, D
h
F(x : w) is not,
in general, a linear function of w and it is usually the case that
(∇
h
F(x))w ,= D
h
F(x : w).
If F

(x

) is ill conditioned or the Lipschitz constant of F

is large, the
difference between them may be significant and a difference approximation to
a directional derivative, which uses the size of w, is likely to be more accurate
than ∇
h
F(x)w.
While we used full matrix approximations to numerical Jacobians in all
the examples reported here, they should be used with caution.
5.4.5. The secant method. The results in this section are for single
equations f(x) = 0, where f is a real-valued function of one real variable.
We assume for the sake of simplicity that there are no errors in the evaluation
of f, i.e., (x) = 0. The standard assumptions, then, imply that f

(x

) ,= 0.
There is no reason to use a difference approximation to f

for equations in a
single variable because one can use previously computed data to approximate
f

(x
n
) by
a
n
=
f(x
n
) −f(x
n−1
)
x
n
−x
n−1
.
The resulting method is called the secant method and is due to Newton
[140], [137]. We will prove that the secant method is locally q-superlinearly
convergent if the standard assumptions hold.
In order to start, one needs two approximations to x

, x
0
and x
−1
. The
local convergence theory can be easily described by Theorem 5.4.1. Let x
c
be
the current iterate and x

the previous iterate. We have (x
c
) = 0 and
∆(x
c
) =
f(x
c
) −f(x

)
x
c
−x

−f

(x
c
). (5.16)
We have the following theorem.
Theorem 5.4.7. Let the standard assumptions hold with N = 1. Then
there is δ > 0 such that if x
0
, x
−1
∈ B(δ) and x
0
,= x
−1
then the secant iterates
converge q-superlinearly to x

.
Proof. Let δ be small enough so that B(δ) ⊂ Ω and the conclusions of
Theorem 5.4.1 hold. Assume that x
−1
, x
0
, . . . , x
n
∈ B(δ). If x
n
= x

for some
finite n, we are done. Otherwise x
n
,= x
n−1
. If we set s = x
n
−x
n−1
we have
by (5.16) and the fundamental theorem of calculus
∆(x
n
) =
_
1
0
(f

(x
n−1
+ts) −f

(x
n
)) dt
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 83
and hence
[∆(x
n
)[ ≤ γ[x
n−1
−x
n
[/2 ≤
γ([e
n
[ +[e
n−1
[)
2
.
Applying Theorem 5.4.1 implies
[e
n+1
[ ≤
¯
K((1 +γ/2)[e
n
[
2
+γ[e
n
[[e
n−1
[/2). (5.17)
If we reduce δ so that
¯
K(1 +γ)δ = η < 1
then x
n
→ x

q-linearly. Therefore
[e
n+1
[
[e
n
[

¯
K((1 +γ/2)[e
n
[ +γ[e
n−1
[/2) → 0
as n → ∞. This completes the proof.
The secant method and the classical bisection (see Exercise 5.7.8) method
are the basis for the popular method of Brent [17] for the solution of single
equations.
5.5. The Kantorovich Theorem
In this section we present the result of Kantorovich [107], [106], which states
that if the standard assumptions “almost” hold at a point x
0
, then there
is a root and the Newton iteration converges r-quadratically to that root.
This result is of use, for example, in proving that discretizations of nonlinear
differential and integral equations have solutions that are near to that of the
continuous problem.
We require the following assumption.
Assumption 5.5.1. There are β, η, ¯ r, and γ with βηγ ≤ 1/2 and x
0
∈ R
N
such that
1. F is differentiable at x
0
,
|F

(x
0
)
−1
| ≤ β, and |F

(x
0
)
−1
F(x
0
)| ≤ η.
2. F

is Lipschitz continuous with Lipschitz constant γ in a ball of radius
¯ r ≥ r

about x
0
where
r

=
1 −

1 −2βηγ
βγ
.
We will let B
0
be the closed ball of radius r

about x
0
B
0
= ¦x[ |x −x
0
| ≤ r

¦.
We do not prove the result in its full generality and refer to [106], [145],
[57], and [58] for a complete proof and for discussions of related results.
However, we give a simplified version of a Kantorovich-like result for the chord
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
84 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
method to show how existence of a solution can be derived from estimates
on the function and Jacobian. Unlike the chord method results in [106],
[145], and [57], Theorem 5.5.1 assumes a bit more than Assumption 5.5.1 but
has a simple proof that uses only the contraction mapping theorem and the
fundamental theorem of calculus and obtains q-linear convergence instead of
r-linear convergence.
Theorem 5.5.1. Let Assumption 5.5.1 hold and let
βγη < 1/2. (5.18)
Then there is a unique root x

of F in B
0
, the chord iteration with x
0
as the
initial iterate converges to x

q-linearly with q-factor βγr

, and x
n
∈ B
0
for
all n. Moreover x

is the unique root of F in the ball of radius
min
_
¯ r,
1 +

1 −2βηγ
βγ
_
about x
0
.
Proof. We will show that the map
φ(x) = x −F

(x
0
)
−1
F(x)
is a contraction on the set
o = ¦x[ |x −x
0
| ≤ r

¦.
By the fundamental theorem of calculus and Assumption 5.5.1 we have for
all x ∈ o
F

(x
0
)
−1
F(x) = F

(x
0
)
−1
F(x
0
)
+F

(x
0
)
−1
_
1
0
F

(x
0
+t(x −x
0
))(x −x
0
) dt
= F

(x
0
)
−1
F(x
0
) +x −x
0
+F

(x
0
)
−1
_
1
0
(F

(x
0
+t(x −x
0
)) −F

(x
0
))(x −x
0
) dt
(5.19)
and
φ(x) −x
0
= −F

(x
0
)
−1
F(x
0
)
−F

(x
0
)
−1
_
1
0
(F

(x
0
+t(x −x
0
)) −F

(x
0
))(x −x
0
) dt.
Hence,
|φ(x) −x
0
| ≤ η +βγr
2

/2 = r

.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 85
and φ maps o into itself.
To show that φ is a contraction on o and to prove the q-linear convergence
we will show that
|φ(x) −φ(y)| ≤ βγr

|x −y|
for all x, y ∈ o. If we do this, the contraction mapping theorem (Theo-
rem 4.2.1) will imply that there is a unique fixed point x

of φ in o and
that x
n
→ x

q-linearly with q-factor βγr

. Clearly x

is a root of F because
it is a fixed point of φ. We know that βγr

< 1 by our assumption that
βηγ < 1/2.
Note that for all x ∈ B
0
φ

(x) = I −F

(x
0
)
−1
F

(x) = F

(x
0
)
−1
(F

(x
0
) −F

(x))
and hence

(x)| ≤ βγ|x −x
0
| ≤ βγr

< 1.
Therefore, for all x, y ∈ B
0
,
|φ(x) −φ(y)| =
_
_
_
_
_
1
0
φ

(y +t(x −y))(x −y) dt
_
_
_
_
≤ βγr

|x −y|.
This proves the convergence result and the uniqueness of x

in B
0
.
To prove the remainder of the uniqueness assertion let
r
+
=
1 +

1 −2βηγ
βγ
and note that r
+
> r

because βηγ < 1/2. If ¯ r = r

, then we are done by the
uniqueness of x

in B
0
. Hence we may assume that ¯ r > r

. We must show
that if x is such that
r

< |x −x
0
| < min(¯ r, r
+
)
then F(x) ,= 0. Letting r = |x −x
0
| we have by (5.19)
|F

(x
0
)
−1
F(x)| ≥ r −η −βγr
2
/2 > 0
because r

< r < r
+
. This completes the proof.
The Kantorovich theorem is more precise than Theorem 5.5.1 and has a
very different proof. We state the theorem in the standard way [145], [106],
[63], using the r-quadratic estimates from [106].
Theorem 5.5.2. Let Assumption 5.5.1 hold. Then there is a unique root
x

of F in B
0
, the Newton iteration with x
0
as the initial iterate converges to
x

, and x
n
∈ B
0
for all n. x

is the unique root of F in the ball of radius
min
_
¯ r,
1 +

1 −2βηγ
βγ
_
about x
0
and the errors satisfy
|e
n
| ≤
(2βηγ)
2
n
2
n
βγ
. (5.20)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
86 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
If βηγ < 1/2 then (5.20) implies that the convergence is r-quadratic (see
Exercise 5.7.16).
5.6. Examples for Newton’s method
In the collection of MATLAB codes we provide an implementation nsol
of Newton’s method, the chord method, the Shamanskii method, or a
combination of all of them based on the reduction in the nonlinear residual.
This latter option decides if the Jacobian should be recomputed based on the
ratio of successive residuals. If
|F(x
+
)|/|F(x
c
)|
is below a given threshold, the Jacobian is not recomputed. If the ratio is too
large, however, we compute and factor F

(x
+
) for use in the subsequent chord
steps. In addition to the inputs of Algorithm sham, a threshold for the ratio
of successive residuals is also input. The Jacobian is recomputed and factored
if either the ratio of successive residuals exceeds the threshold 0 < ρ < 1 or
the number of iterations without an update exceeds m. The output of the
MATLAB implementation of nsol is the solution, the history of the iteration
stored as the vector of l

norms of the nonlinear residuals, and an error flag.
nsol uses diffjac to approximate the Jacobian by ∇
h
F with h = 10
−7
.
While our examples have dense Jacobians, the ideas in the implementation
of nsol also apply to situations in which the Jacobian is sparse and direct
methods for solving the equation for the Newton step are necessary. One would
still want to minimize the number of Jacobian computations (by the methods
of [47] or [42], say) and sparse matrix factorizations. In Exercise 5.7.25, which
is not easy, the reader is asked to create a sparse form of nsol.
Algorithm 5.6.1. nsol(x, F, τ, m, ρ)
1. r
c
= r
0
= |F(x)|
2. Do while |F(x)| > τ
r
r
0

a
(a) Compute ∇
h
F(x) ≈ F

(x)
(b) Factor ∇
h
F(x) = LU.
(c) i
s
= 0, σ = 0
(d) Do While i
s
< m and σ ≤ ρ
i. i
s
= i
s
+ 1
ii. Solve LUs = −F(x)
iii. x = x +s
iv. Evaluate F(x)
v. r
+
= |F(x)|, σ = r
+
/r
c
, r
c
= r
+
vi. If |F(x)| ≤ τ
r
r
0

a
exit.
In the MATLAB code the iteration is terminated with an error condition
if σ ≥ 1 at any point. This indicates that the local convergence analysis in this
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 87
section is not applicable. In Chapter 8 we discuss how to recover from this.
nsol is based on the assumption that a difference Jacobian will be used. In
Exercise 5.7.21 you are asked to modify the code to accept an analytic Jacobian.
Many times the Jacobian can be computed efficiently by reusing results that
are already available from the function evaluation. The parameters m and ρ
are assigned the default values of 1000 and .5. With these default parameters
the decision to update the Jacobian is made based entirely on the reduction in
the norm of the nonlinear residual. nsol allows the user to specify a maximum
number of iterations; the default is 40.
The Chandrasekhar H-equation. The Chandrasekhar H-equation, [41],
[30],
F(H)(µ) = H(µ) −
_
1 −
c
2
_
1
0
µH(ν) dν
µ +ν
_
−1
= 0, (5.21)
is used to solve exit distribution problems in radiative transfer.
We will discretize the equation with the composite midpoint rule. Here we
approximate integrals on [0, 1] by
_
1
0
f(µ) dµ ≈
1
N
N

j=1
f(µ
j
),
where µ
i
= (i −1/2)/N for 1 ≤ i ≤ N. The resulting discrete problem is
F(x)
i
= x
i

_
_
1 −
c
2N
N

j=1
µ
i
x
j
µ
i

j
_
_
−1
. (5.22)
It is known [132] that both (5.21) and the discrete analog (5.22) have two
solutions for c ∈ (0, 1). Only one of these solutions has physical meaning,
however. Newton’s method, with the 0 function or the constant function with
value 1 as initial iterate, will find the physically meaningful solution [110].
The standard assumptions hold near either solution [110] for c ∈ [0, 1). The
discrete problem has its own physical meaning [41] and we will attempt to
solve it to high accuracy. Because H has a singularity at µ = 0, the solution of
the discrete problem is not even a first-order accurate approximation to that
of the continuous problem.
In the computations reported here we set N = 100 and c = .9. We used the
function identically one as initial iterate. We computed all Jacobians with the
MATLAB code diffjac. We set the termination parameters τ
r
= τ
a
= 10
−6
.
We begin with a comparison of Newton’s method and the chord method.
We present some iteration statistics in both tabular and graphical form. In
Table 5.1 we tabulate the iteration counter n, |F(x
n
)|

/|F(x
0
)|

, and the
ratio
R
n
= |F(x
n
)|

/|F(x
n−1
)|

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
88 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Table 5.1
Comparison of Newton and chord iteration.
n |F(x
n
)|

/|F(x
0
)|

R
n
|F(x
n
)|

/|F(x
0
)|

R
n
0 1.000e+00 1.000e+00
1 1.480e-01 1.480e-01 1.480e-01 1.480e-01
2 2.698e-03 1.823e-02 3.074e-02 2.077e-01
3 7.729e-07 2.865e-04 6.511e-03 2.118e-01
4 1.388e-03 2.132e-01
5 2.965e-04 2.136e-01
6 6.334e-05 2.136e-01
7 1.353e-05 2.136e-01
8 2.891e-06 2.136e-01
for n ≥ 1 for both Newton’s method and the chord method. In Fig. 5.1
the solid curve is a plot of |F(x
n
)|

/|F(x
0
)|

for Newton’s method and
the dashed curve a plot of |F(x
n
)|

/|F(x
0
)|

for the chord method. The
concave iteration track is the signature of superlinear convergence, the linear
track indicates linear convergence. See Exercise 5.7.15 for a more careful
examination of this concavity. The linear convergence of the chord method
can also be seen in the convergence of the ratios R
n
to a constant value.
For this example the MATLAB flops command returns an fairly accurate
indicator of the cost. The Newton iteration required over 8.8 million floating-
point operations while the chord iteration required under 3.3 million. In
Exercise 5.7.18 you are asked to obtain timings in your environment and will
see how the floating-point operation counts are related to the actual timings.
The good performance of the chord method is no accident. If we think
of the chord method as a preconditioned nonlinear Richardson iteration with
the initial Jacobian playing the role of preconditioner, then the chord method
should be much more efficient than Newton’s method if the equations for the
steps can be solved cheaply. In the case considered here, the cost of a single
evaluation and factorization of the Jacobian is far more than the cost of the
entire remainder of the chord iteration.
The ideas in Chapters 6 and 7 show how the low iteration cost of the
chord method can be combined with the small number of iterations of a
superlinearly convergent nonlinear iterative method. In the context of this
particular example, the cost of the Jacobian computation can be reduced by
analytic evaluation and reuse of data that have already been computed for the
evaluation of F. The reader is encouraged to do this in Exercise 5.7.21.
In the remaining examples, we plot, but do not tabulate, the iteration
statistics. In Fig. 5.2 we compare the Shamanskii method with m = 2 with
Newton’s method. The Shamanskii computation required under 6 million
floating-point operations, as compared with over 8.8 for Newton’s method.
However, for this problem, the simple chord method is the most efficient.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 89
0 1 2 3 4 5 6 7 8
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 5.1. Newton and chord methods for H-equation, c = .9.
0 0.5 1 1.5 2 2.5 3 3.5 4
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 5.2. Newton and Shamanskii method for H-equation, c = .9.
The hybrid approach in nsol with the default parameters would be the chord
method for this problem.
We now reconsider the H-equation with c = .9999. For this problem the
Jacobian at the solution is nearly singular [110]. Aside from c, all iteration
parameters are the same as in the previous example. While Newton’s method
converges in 7 iterations and requires 21 million floating-point operations, the
chord method requires 188 iterations and 10.6 million floating-point operations.
The q-factor for the chord method in this computation was over .96, indicating
very slow convergence. The hybrid method in nsol with the default parameters
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
90 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
computes the Jacobian four times during the course of the iteration, converges
in 14 iterations, and requires 12 million floating-point operations. In Fig. 5.3
we compare the Newton iteration with the hybrid.
0 2 4 6 8 10 12 14
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 5.3. Newton and hybrid methods for H-equation c = .9999.
One should approach plots of convergence histories with some caution. The
theory for the chord method asserts that the error norms converge q-linearly to
zero. This implies only that the nonlinear residual norms ¦|F(x
n
)|¦ converge
r-linearly to zero. While the plots indicate q-linear convergence of the nonlinear
residuals, the theory does not, in general, support this. In practice, however,
plots like those in this section are common.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 91
5.7. Exercises on Newton’s method
In some of the exercises here, and in the rest of this part of the book, you will
be asked to plot or tabulate iteration statistics. When asked to do this, for
each iteration tabulate the iteration counter, the norm of F, and the ratio
of |F(x
n
)|/|F(x
n−1
)| for n ≥ 1. A better alternative in the MATLAB
environment is to use the semilogy command to plot the norms. When
one does this, one can visually inspect the plot to determine superlinear
convergence (concavity) without explicitly computing the ratios. Use the l

norm.
5.7.1. A function G is said to be H¨older continuous with exponent α in Ω if
|G(x) −G(y)| ≤ K|x −y|
α
for all x, y ∈ Ω. Show that if the Lipschitz
continuity condition on F

in the standard assumptions is replaced by
H¨older continuity with exponent α > 0 that the Newton iterates converge
with q-order 1 +α.
5.7.2. Can the performance of the Newton iteration be improved by a linear
change of variables? That is, for nonsingular N N matrices A and
B, can the Newton iterates for F(x) = 0 and AF(Bx) = 0 show any
performance difference when started at the same initial iterate? What
about the chord method?
5.7.3. Let ∇
h
F be given by Definition 5.4.1. Given the standard assumptions,
prove that the iteration given by
x
n+1
= x
n
−(∇
hn
F(x
n
))
−1
F(x
n
)
is locally q-superlinearly convergent if h
n
→ 0. When is it locally q-
quadratically convergent?
5.7.4. Suppose F

(x
n
) is replaced by ∇
hn
F(x
n
) in the Shamanskii method.
Discuss how h
n
must be decreased as the iteration progresses to preserve
the q-order of m+ 1 [174].
5.7.5. In what way is the convergence analysis of the secant method changed if
there are errors in the evaluation of f?
5.7.6. Prove that the secant method converges r-superlinearly with r-order
(

5 + 1)/2. This is easy.
5.7.7. Show that the secant method converges q-superlinearly with q-order
(

5 + 1)/2.
5.7.8. The bisection method produces a sequence of intervals (x
k
l
, x
k
r
) that
contain a root of a function of one variable. Given (x
k
l
, x
k
r
) with
f(x
k
l
)f(x
k
r
) < 0 we replace one of the endpoints with y = (x
k
l
+ x
k
r
)/2
to force f(x
k+1
l
)f(x
k+1
r
) < 0. If f(y) = 0, we terminate the iteration.
Show that the sequence x
k
= (x
k
l
+ x
k
r
)/2 is r-linearly convergent if f is
continuous and there exists (x
0
l
, x
0
r
) such that f(x
0
l
)f(x
0
r
) < 0.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
92 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
5.7.9. Write a program that solves single nonlinear equations with Newton’s
method, the chord method, and the secant method. For the secant
method, use x
−1
= .99x
0
. Apply your program to the following
function/initial iterate combinations, tabulate or plot iteration statistics,
and explain your results.
1. f(x) = cos(x) −x, x
0
= .5
2. f(x) = tan
−1
(x), x
0
= 1
3. f(x) = sin(x), x
0
= 3
4. f(x) = x
2
, x
0
= .5
5. f(x) = x
2
+ 1, x
0
= 10
5.7.10. For scalar functions of one variable a quadratic model of f about a point
x
c
is
m
c
(x) = f(x
c
) +f

(x
c
)(x −x
c
) +f

(x
c
)(x −x
c
)
2
/2.
Consider the iterative method that defines x
+
as a root of m
c
. What
problems might arise in such a method? Resolve these problems and
implement your algorithm. Test it on the problems in Exercise 5.7.9. In
what cases is this algorithm superior to Newton’s method? For hints and
generalizations of this idea to functions of several variables, see [170]. Can
the ideas in [170] be applied to some of the high-order methods discussed
in [190] for equations in one unknown?
5.7.11. Show that if f is a real function of one real variable, f

is Lipschitz
continuous, and f(x

) = f

(x

) = 0 but f

(x

) ,= 0 then the iteration
x
n+1
= x
n
−2f(x
n
)/f

(x
n
)
converges locally q-quadratically to x

provided x
0
is sufficiently near to
x

, but not equal to x

[171].
5.7.12. Show that if f is a real function of one real variable, the standard
assumptions hold, f

(x

) = 0 and f

(x) is Lipschitz continuous, then
the Newton iteration converges cubically (q-superlinear with q-order 3).
5.7.13. Show that if f is a real function of one real variable, the standard
assumptions hold, x
0
is sufficiently near x

, and f(x
0
)f

(x
0
) > 0, then
the Newton iteration converges monotonically to x

. What if f

(x

) = 0?
How can you extend this result if f

(x

) = f

(x

) = . . . f
(k)
(x

) = 0 ,=
f
(k+1)
(x

)? See [131].
5.7.14. How would you estimate the q-order of a q-superlinearly convergent
sequence? Rigorously justify your answer.
5.7.15. If x
n
→ x

q-superlinearly with q-order α > 1, show that log(|e
n
|) is a
concave function of n in the sense that
log(|e
n+1
|) −log(|e
n
|)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
NEWTON’S METHOD 93
is a decreasing function of n for n sufficiently large. Is this still the case
if the convergence is q-superlinear but there is no q-order? What if the
convergence is q-linear?
5.7.16. Show that the sequence on the right side of (5.20) is r-quadratically (but
not q-quadratically) convergent to 0 if βηγ < 1/2.
5.7.17. Typically (see [89], [184]) the computed LU factorization of F

(x) is the
exact LU factorization of a nearby matrix.
LU = F

(x) +E.
Use the theory in [89], [184] or another book on numerical linear algebra
and Theorem 5.4.1 to describe how this factorization error affects the
convergence of the Newton iteration. How does the analysis for the QR
factorization differ?
5.7.18. Duplicate the results on the H-equation in ¸ 5.6. Try different values of
N and c. How does the performance of the methods differ as N and c
change? Compare your results to the tabulated values of the H function
in [41] or [14]. Compare execution times in your environment. Tabulate
or plot iteration statistics. Use the MATLAB cputime command to see
how long the entire iteration took for each N, c combination and the
MATLAB flops command to get an idea for the number of floating-
point operations required by the various methods.
5.7.19. Solve the H-equation with c = 1 and N = 100. Explain your results
(see [50] and [157]). Does the trick in Exercise 5.7.11 improve the
convergence? Try one of the approaches in [51], [49], or [118] for
acceleration of the convergence. You might also look at [90], [170], [75],
[35] and [155] for more discussion of this. Try the chord method and
compare the results to part 4 of Exercise 5.7.9 (see [52]).
5.7.20. Solve the H-equation with N = 100, c = .9 and c = .9999 with fixed-point
iteration. Compare timings and flop counts with Newton’s method, the
chord method, and the algorithm in nsol.
5.7.21. Compute by hand the Jacobian of F, the discretized nonlinearity from
the H equation in (5.22). Prove that F

can be computed at a cost of
less than twice that of an evaluation of F. Modify nsol to accept this
analytic Jacobian rather than evaluate a difference Jacobian. How do
the execution times and flop counts change?
5.7.22. Modify dirder by setting the forward difference step h = 10
−2
. How
does this change affect the convergence rate observations that you have
made?
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
94 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
5.7.23. Add noise to your function evaluation with a random number generator,
such as the the MATLAB rand function. If the size of the noise is
O(10
−4
) and you make the appropriate adjustments to dirder, how are
the convergence rates different? What is an appropriate termination
criterion? At what point does the iteration stagnate?
5.7.24. Assume that the standard assumptions hold, that the cost of a function
evaluation is O(N
2
) floating-point operations, the cost of a Jacobian is
O(N) function evaluations, and that x
0
is near enough to x

so that the
Newton iteration converges q-quadratically to x

. Show that the number
n of Newton iterations needed to obtain |e
n
| ≤ |e
0
| is O(log()) and
that the number of floating-point operations required is O(N
3
log()).
What about the chord method?
5.7.25. Develop a sparse version of nsol for problems with tridiagonal Jacobian.
In MATLAB, for example, you could modify diffjac to use the
techniques of [47] and nsol to use the sparse matrix factorization in
MATLAB. Apply your code to the central difference discretization of the
two-point boundary value problem
−u

= sin(u) +f(x), u(0) = u(1) = 0.
Here, the solution is u

(x) = x(1 −x), and
f(x) = 2 −sin(x(1 −x)).
Let u
0
= 0 be the initial iterate.
5.7.26. Suppose you try to find an eigenvector-eigenvalue pair (φ, λ) of an NN
matrix A by solving the system of N + 1 nonlinear equations
F(φ, λ) =
_
Aφ −λφ
φ
T
φ −1
_
=
_
0
0
_
(5.23)
for the vector x = (φ
T
, λ)
T
∈ R
N+1
. What is the Jacobian of this
system? If x = (φ, λ) ∈ R
N+1
is an eigenvector-eigenvalue pair, when is
F

(x) nonsingular? Relate the application of Newton’s method to (5.23)
to the inverse power method. See [150] for more development of this
idea.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 6
Inexact Newton Methods
Theorem 5.4.1 describes how errors in the derivative/function affect the
progress in the Newton iteration. Another way to look at this is to ask how
an approximate solution of the linear equation for the Newton step affects the
iteration. This was the view taken in [55] where inexact Newton methods in
which the step satisfies
|F

(x
c
)s +F(x
c
)| ≤ η
c
|F(x
c
)| (6.1)
are considered. Any approximate step is accepted provided that the relative
residual of the linear equation is small. This is quite useful because conditions
like (6.1) are precisely the small linear residual termination conditions for
iterative solution of the linear system for the Newton step. Such methods are
not new. See [145] and [175], for example, for discussion of these ideas in the
context of the classical stationary iterative methods. In most of this chapter
we will focus on the approximate solution of the equation for the Newton step
by GMRES, but the other Krylov subspace methods discussed in Chapters 2
and 3 and elsewhere can also be used. We follow [69] and refer to the term η
c
on the right hand side of (6.1) as the forcing term.
6.1. The basic estimates
We will discuss specific implementation issues in ¸ 6.2. Before that we will
give the basic result from [55] to show how a step that satisfies (6.1) affects
the convergence. We will present this result in two parts. In ¸ 6.1.1 we give a
straightforward analysis that uses the techniques of Chapter 5. In ¸ 6.1.2 we
show how the requirements of the simple analysis can be changed to admit a
more aggressive (i.e., larger) choice of the parameter η.
6.1.1. Direct analysis. The proof and application of Theorem 6.1.1 should
be compared to that of Theorem 5.1.1 and the other results in Chapter 5.
Theorem 6.1.1. Let the standard assumptions hold. Then there are δ and
K
I
such that if x
c
∈ B(δ), s satisfies (6.1), and x
+
= x
c
+s then
|e
+
| ≤ K
I
(|e
c
| +η
c
)|e
c
|. (6.2)
95
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
96 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Proof. Let δ be small enough so that the conclusions of Lemma 4.3.1 and
Theorem 5.1.1 hold. To prove the first assertion (6.2) note that if
2
r = −F

(x
c
)s −F(x
c
)
is the linear residual then
s +F

(x
c
)
−1
F(x
c
) = −F

(x
c
)
−1
r
and
e
+
= e
c
+s = e
c
−F

(x
c
)
−1
F(x
c
) −F

(x
c
)
−1
r. (6.3)
Now, (6.1), (4.7), and (4.6) imply that
|s +F

(x
c
)
−1
F(x
c
)| ≤ |F

(x
c
)
−1

c
|F(x
c
)|
≤ 4κ(F

(x

))η
c
|e
c
|.
Hence, using (6.3) and Theorem 5.1.1
|e
+
| ≤ |e
c
−F

(x
c
)
−1
F(x
c
)| + 4κ(F

(x

))η
c
|e
c
|
≤ K|e
c
|
2
+ 4κ(F

(x

))η
c
|e
c
|,
where K is the constant from (5.2). If we set
K
I
= K + 4κ(F

(x

)),
the proof is complete.
We summarize the implications of Theorem 6.1.1 for iterative methods in
the next result, which is a direct consequence of (6.2) and will suffice to explain
most computational observations.
Theorem 6.1.2. Let the standard assumptions hold. Then there are δ and
¯ η such that if x
0
∈ B(δ), ¦η
n
¦ ⊂ [0, ¯ η], then the inexact Newton iteration
x
n+1
= x
n
+s
n
,
where
|F

(x
n
)s
n
+F(x
n
)| ≤ η
n
|F(x
n
)|
converges q-linearly to x

. Moreover
• if η
n
→ 0 the convergence is q-superlinear, and
• if η
n
≤ K
η
|F(x
n
)|
p
for some K
η
> 0 the convergence is q-superlinear
with q-order 1 +p.
2
Our definition of r is consistent with the idea that the residual in a linear equation is b −Ax. In
[55] r = F

(xc)s +F(xc).
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 97
Proof. Let δ be small enough so that (6.2) holds for x
c
∈ B(δ). Reduce δ
and ¯ η if needed so that
K
I
(δ + ¯ η) < 1,
where K
I
is from (6.2). Then if n ≥ 0 and x
n
∈ B(δ) we have
|e
n+1
| ≤ K
I
(|e
n
| +η
n
)|e
n
| ≤ K
I
(δ + ¯ η)|e
n
| < |e
n
|.
This proves q-linear convergence with a q-factor of K
I
(δ + ¯ η).
If η
n
→ 0 then q-superlinear convergence follows from the definition. If
η
n
≤ K
η
|F(x
n
)|
p
then we may use (4.7) and (6.2) to conclude
|e
n+1
| ≤ K
I
(|e
n
|
1−p
+K
η
2
p
|F

(x

)|
p
)|e
n
|
1+p
which completes the proof.
6.1.2. Weighted norm analysis. Since K
I
= O(κ(F

(x

))), one might
conclude from Theorem 6.1.2 that if F

(x

) is ill conditioned very small forcing
terms must be used. This is not the case and the purpose of this subsection
is to describe more accurately the necessary restrictions on the forcing terms.
The results here differ from those in ¸ 6.1.1 in that no restriction is put on
the sequence ¦η
n
¦ ⊂ [0, 1) other than requiring that 1 not be an accumulation
point.
Theorem 6.1.3. Let the standard assumptions hold. Then there is δ such
that if x
c
∈ B(δ), s satisfies (6.1), x
+
= x
c
+s, and η
c
≤ η < ¯ η < 1, then
|F

(x

)e
+
| ≤ ¯ η|F

(x

)e
c
|. (6.4)
Proof. To prove (6.4) note that Theorem 4.0.1 implies that
|F(x
c
)| ≤ |F

(x

)e
c
| +
γ|e
c
|
2
2
.
Since
|e
c
| = |F

(x

)
−1
F

(x

)e
c
| ≤ |F

(x

)
−1
||F

(x

)e
c
|
we have, with
M
0
=
γ|F

(x

)
−1
|
2
|F(x
c
)| ≤ (1 +M
0
δ)|F

(x

)e
c
|. (6.5)
Now,
F

(x

)e
+
= F

(x

)(e
c
+s)
= F

(x

)(e
c
−F

(x
c
)
−1
F(x
c
) −F

(x
c
)
−1
r).
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
98 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
By Theorem 5.1.1
|F

(x

)(e
c
−F

(x
c
)
−1
F(x
c
))| ≤ K|F

(x

)||e
c
|
2
.
Hence,
|F

(x

)e
+
| ≤ |F

(x

)F

(x
c
)
−1
r| +K|F

(x

)||e
c
|
2
. (6.6)
Since
|F

(x

)F

(x
c
)
−1
r| ≤ |r| +|(F

(x

) −F

(x
c
))F

(x
c
)
−1
r|
≤ (1 + 2γ|F

(x

)
−1
||e
c
|)|r|,
we may set
M
1
= 2γ|F

(x

)
−1
|, and M
2
= K|F

(x

)|
and obtain, using (4.7) and (6.5),
|F

(x

)e
+
| ≤ (1 +M
1
δ)|r| +M
2
δ|e
c
|
≤ (1 +M
1
δ)(1 +M
0
δ)η
c
|F

(x

)e
c
|
+M
2
δ|F

(x

)
−1
||F

(x

)e
c
|
≤ ((1 +M
1
δ)(1 +M
0
δ)η +M
2
δ|F

(x

)
−1
|)|F

(x

)e
c
|.
(6.7)
Now let δ be small enough so that
(1 +M
1
δ)(1 +M
0
δ)η +M
2
δ|F

(x

)
−1
| ≤ ¯ η
and the proof is complete.
Note that the distance δ from the initial iterate to the solution may have
to be smaller for (6.4) to hold than for (6.2). However (6.4) is remarkable in its
assertion that any method that produces a step with a linear residual less than
that of the zero step will reduce the norm of the error if the error is measured
with the weighted norm
| |

= |F

(x

) |.
In fact, Theorem 6.1.3 asserts q-linear convergence in the weighted norm if the
initial iterate is sufficiently near x

. This is made precise in Theorem 6.1.4.
The application of Theorem 6.1.3 described in Theorem 6.1.4 differs from
Theorem 6.1.2 in that we do not demand that ¦η
n
¦ be bounded away from 1
by a sufficiently large amount, just that 1 not be an accumulation point. The
importance of this theorem for implementation is that choices of the sequence
of forcing terms ¦η
n
¦ (such as η
n
= .5 for all n) that try to minimize the
number of inner iterations are completely justified by this result if the initial
iterate is sufficiently near the solution. We make such a choice in one of the
examples considered in ¸ 6.4 and compare it to a modification of a choice from
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 99
[69] that decreases η
n
rapidly with a view toward minimizing the number of
outer iterations.
Theorem 6.1.4. Let the standard assumptions hold. Then there is δ such
that if x
0
∈ B(δ), ¦η
n
¦ ⊂ [0, η] with η < ¯ η < 1, then the inexact Newton
iteration
x
n+1
= x
n
+s
n
,
where
|F

(x
n
)s
n
+F(x
n
)| ≤ η
n
|F(x
n
)|
converges q-linearly with respect to | |

to x

. Moreover
• if η
n
→ 0 the convergence is q-superlinear, and
• if η
n
≤ K
η
|F(x
n
)|
p
for some K
η
> 0 the convergence is q-superlinear
with q-order 1 +p.
Proof. The proof uses both (6.4) and (6.2). Our first task is to relate the
norms | | and | |

so that we can estimate δ. Let δ
0
be such that (6.4) holds
for |e
c
| < δ
0
.
For all x ∈ R
N
,
|x|

≤ |F

(x

)||x| ≤ κ(F

(x

))|x|

, (6.8)
Note that |e| < δ
0
if
|e|

< δ

= |F

(x

)|δ
0
Set δ = |F

(x

)|
−1
δ

. Then |e
0
|

< δ

if |e
0
| < δ.
Our proof does not rely on the (possibly false) assertions that |e
n
| < δ for
all n or that x
n
→ x

q-linearly with respect to the unweighted norm. Rather
we note that (6.4) implies that if |e
n
|

< δ

then
|e
n+1
|

≤ ¯ η|e
n
|

< δ

(6.9)
and hence x
n
→ x

q-linearly with respect to the weighted norm. This proves
the first assertion.
To prove the assertions on q-superlinear convergence, note that since
x
n
→ x

, eventually |e
n
| will be small enough so that the conclusions of
Theorem 6.1.1 hold. The superlinear convergence results then follow exactly
as in the proof of Theorem 6.1.2.
We close this section with some remarks on the two types of results in
Theorems 6.1.2 and 6.1.4. The most significant difference is the norm in which
q-linear convergence takes place. q-linear convergence with respect to the
weighted norm | |

is equivalent to q-linear convergence of the sequence of
nonlinear residuals ¦|F(x
n
)|¦. We state this result as Proposition 6.1.1 and
leave the proof to the reader in Exercise 6.5.2. The implications for the choice of
the forcing terms ¦η
n
¦ are also different. While Theorem 6.1.2 might lead one
to believe that a small value of η
n
is necessary for convergence, Theorem 6.1.4
shows that the sequence of forcing terms need only be kept bounded away from
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
100 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
1. A constant sequence such as η
n
= .5 may well suffice for linear convergence,
but at a price of a greater number of outer iterations. We will discuss other
factors in the choice of forcing terms in ¸ 6.3.
Proposition 6.1.1. Let the standard assumptions hold and let x
n
→ x

.
Then |F(x
n
)| converges q-linearly to 0 if and only if |e
n
|

does.
6.1.3. Errors in the function. In this section, we state two theorems on
inexact local improvement. We leave the proofs, which are direct analogs to
the proofs of Theorems 6.1.1 and 6.1.3, to the reader as exercises. Note that
errors in the derivative, such as those arising from a difference approximation
of the action of F

(x
c
) on a vector, can be regarded as part of the inexact
solution of the equation for the Newton step. See [55] or Proposition 6.2.1 and
its proof for an illustration of this point.
Theorem 6.1.5. Let the standard assumptions hold. Then there are δ and
K
I
such that if x
c
∈ B(δ), s satisfies
|F

(x
c
)s +F(x
c
) +(x
c
)| ≤ η
c
|F(x
c
) +(x
c
)| (6.10)
and x
+
= x
c
+s, then
|e
+
| ≤ K
I
((|e
c
| +η
c
)|e
c
| +|(x
c
)|). (6.11)
Theorem 6.1.6. Let the standard assumptions hold. Then there is δ such
that if x
c
∈ B(δ), s satisfies (6.10), x
+
= x
c
+ s, and η
c
≤ η < ¯ η < 1, then
there is K
E
such that
|F

(x

)e
+
| ≤ ¯ η|F

(x

)e
c
| +K
E
|(x
c
)|. (6.12)
6.2. Newton-iterative methods
A Newton-iterative method realizes (6.1) with an iterative method for the linear
system for the Newton step, terminating when the relative linear residual is
smaller than η
c
(i.e, when (6.1) holds). The name of the method indicates
which linear iterative method is used, for example, Newton-SOR, Newton-CG,
Newton-GMRES, would use SOR, CG, or GMRES to solve the linear system.
This naming convention is taken from [175] and [145]. These methods have also
been called truncated Newton methods [56], [136] in the context of optimization.
Typically the nonlinear iteration that generates the sequence ¦x
n
¦ is called
the outer iteration and the linear iteration that generates the approximations
to the steps is called the inner iteration.
In this section we use the l
2
norm to measure nonlinear residuals. The
reason for this is that the Krylov methods use the scalar product and the
estimates in Chapters 2 and 3 are in terms of the Euclidean norm. In the
examples discussed in ¸ 6.4 we will scale the l
2
norm by a factor of 1/N so
that the results for the differential and integral equations will be independent
of the computational mesh.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 101
6.2.1. Newton GMRES. We provide a detailed analysis of the Newton-
GMRES iteration. We begin by discussing the effects of a forward difference
approximation to the action of the Jacobian on a vector.
If the linear iterative method is any one of the Krylov subspace methods
discussed in Chapters 2 and 3 then each inner iteration requires at least one
evaluation of the action of F

(x
c
) on a vector. In the case of CGNR and CGNE,
an evaluation of the action of F

(x
c
)
T
on a vector is also required. In many
implementations [20], [24], the action of F

(x
c
) on a vector w is approximated
by a forward difference, (5.15), D
h
F(x : w) for some h. It is important to note
that this is entirely different from forming the finite difference Jacobian ∇
h
F(x)
and applying that matrix to w. In fact, as pointed out in [20], application
of GMRES to the linear equation for the Newton step with matrix-vector
products approximated by finite differences is the same as the application of
GMRES to the matrix G
h
F(x) whose last k −1 columns are the vectors
v
k
= D
h
F(x : v
k−1
)
The sequence ¦v
k
¦ is formed in the course of the forward-difference GMRES
iteration.
To illustrate this point, we give the forward-difference GMRES algorithm
for computation of the Newton step, s = −F

(x)
−1
F(x). Note that matrix-
vector products are replaced by forward difference approximations to F

(x), a
sequence of approximate steps ¦s
k
¦ is produced, that b = −F(x), and that the
initial iterate for the linear problem is the zero vector. We give a MATLAB
implementation of this algorithm in the collection of MATLAB codes.
Algorithm 6.2.1. fdgmres(s, x, F, h, η, kmax, ρ)
1. s = 0, r = −F(x), v
1
= r/|r|
2
, ρ = |r|
2
, β = ρ, k = 0
2. While ρ > η|F(x)|
2
and k < kmax do
(a) k = k + 1
(b) v
k+1
= D
h
F(x : v
k
)
for j = 1, . . . k
i. h
jk
= v
T
k+1
v
j
ii. v
k+1
= v
k+1
−h
jk
v
j
(c) h
k+1,k
= |v
k+1
|
2
(d) v
k+1
= v
k+1
/|v
k+1
|
2
(e) e
1
= (1, 0, . . . , 0)
T
∈ R
k+1
Minimize |βe
1
−H
k
y
k
|
R
k+1 to obtain y
k
∈ R
k
.
(f) ρ = |βe
1
−H
k
y
k
|
R
k+1.
3. s = V
k
y
k
.
In our MATLAB implementation we solve the least squares problem in
step 2e by the Givens rotation approach used in Algorithm gmres.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
102 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
As should be clear from inspection of Algorithm gmres, the difference
between ∇
h
F and G
h
F is that the columns are computed by taking directional
derivatives based on two different orthonormal bases: ∇
h
F using the basis of
unit vectors in coordinate directions and G
h
F the basis for the Krylov space /
k
constructed by algorithm fdgmres. The action of G
h
F on /

k
, the orthogonal
complement of /
k
, is not used by the algorithm and, for the purposes of
analysis, could be specified by the action of ∇
h
F on any basis for /

k
. Hence
G
h
F is also first order accurate in h. Assuming that there is no error in the
evaluation of F we may summarize the observations above in the following
proposition, which is a special case of the more general results in [20].
Proposition 6.2.1. Let the standard assumptions hold. Let η ∈ (0, 1).
Then there are C
G
,
¯
h, and δ such that if x ∈ B(δ), h ≤
¯
h, and Algo-
rithm fdgmres terminates with k < kmax then the computed step s satisfies
|F

(x)s +F(x)|
2
< (η +C
G
h)|F(x)|
2
. (6.13)
Proof. First let δ be small enough so that B(δ) ⊂ Ω and the conclusions
to Lemma 4.3.1 hold. Let ¦u
j
¦ be any orthonormal basis for R
N
such that
u
j
= v
j
for 1 ≤ j ≤ k, where ¦v
j
¦
k
j=1
is the orthonormal basis for /
k
generated
by Algorithm fdgmres.
Consider the linear map defined by its action on ¦u
j
¦
Bu
j
= D
h
F(x : u
j
).
Note that
D
h
F(x : u
j
) =
_
1
0
F

(x +th|x|
2
u
j
)u
j
dt
= F

(x)u
j
+
_
1
0
(F

(x +th|x|
2
u
j
) −F

(x))u
j
dt.
Since F

is Lipschitz continuous and ¦u
j
¦ is an orthonormal basis for R
N
we
have, with ¯ γ = γ(|x

|
2
+δ),
|B −F

(x)|
2
≤ h|x|
2
γ/2 ≤ h¯ γ/2 (6.14)
Since the linear iteration (fdgmres) terminates in k < kmax iterations, we
have, since B and G
h
F agree on /
k
,
|Bs +F(x)|
2
≤ η|F(x)|
2
and therefore
|F

(x)s +F(x)|
2
≤ η|F(x)|
2
+h¯ γ|s|
2
/2. (6.15)
Assume that
¯
h¯ γ ≤ |F

(x

)
−1
|
−1
2
/2
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 103
Then Lemma 4.3.1 and (6.15) imply
|F

(x

)
−1
|
−1
2
|s|
2
/2 ≤ |F

(x)
−1
|
−1
2
|s|
2
≤ |F

(x)s|
2
≤ (1 +η)|F(x)|
2
+
¯
h¯ γ|s|
2
/2.
Therefore,
|s|
2
≤ 4(1 +η)|F

(x

)
−1
|
2
|F(x)|
2
. (6.16)
Combining (6.16) with (6.15) completes the proof with C
G
= 4¯ γ(1 +
η)|F

(x

)
−1
|
2
.
Proposition 6.2.1 implies that a finite difference implementation will not
affect the performance of Newton-GMRES if the steps in the forward difference
approximation of the derivatives are sufficiently small. In an implementation,
therefore, we must specify not only the sequence ¦η
n
¦, but also the steps ¦h
n
¦
used to compute forward differences. Note also that Proposition 6.2.1 applies
to a restarted GMRES because upon successful termination (6.15) will hold.
We summarize our results so far in the analogs of Theorem 6.1.4 and
Theorem 6.1.2. The proofs are immediate consequences of Theorems 6.1.2,
6.1.4, and Proposition 6.2.1 and are left as (easy) exercises.
Theorem 6.2.1. Assume that the assumptions of Proposition 6.2.1 hold.
Then there are δ, ¯ σ such that if x
0
∈ B(δ) and the sequences ¦η
n
¦ and ¦h
n
¦
satisfy
σ
n
= η
n
+C
G
h
n
≤ ¯ σ
then the forward difference Newton-GMRES iteration
x
n+1
= x
n
+s
n
where s
n
is computed by Algorithm fdgmres with arguments
(s
n
, x
n
, F, h
n
, η
n
, kmax, ρ)
converges q-linearly and s
n
satisfies
|F

(x
n
)s
n
+F(x
n
)|
2
< σ
n
|F(x
n
)|
2
. (6.17)
Moreover,
• if σ
n
→ 0 the convergence is q-superlinear, and
• if σ
n
≤ K
η
|F(x
n
)|
p
2
for some K
η
> 0 the convergence is q-superlinear
with q-order 1 +p.
Theorem 6.2.2. Assume that the assumptions of Proposition 6.2.1 hold.
Then there is δ such that if x
0
∈ B(δ) and the sequences ¦η
n
¦ and ¦h
n
¦ satisfy
σ
n
= η
n
+C
G
h
n
∈ [0, ¯ σ]
with 0 < ¯ σ < 1 then the forward difference Newton-GMRES iteration
x
n+1
= x
n
+s
n
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
104 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
where s
n
is computed by Algorithm fdgmres with arguments
(s
n
, x
n
, F, h
n
, η
n
, kmax, ρ)
converges q-linearly with respect to | |

and s
n
satisfies (6.17). Moreover
• if σ
n
→ 0 the convergence is q-superlinear, and
• if σ
n
≤ K
η
|F(x
n
)|
p
2
for some K
η
> 0 the convergence is q-superlinear
with q-order 1 +p.
In the case where h
n
is approximately the square root of machine roundoff,
σ
n
≈ η
n
if η
n
is not too small and Theorem 6.2.2 states that the observed
behavior of the forward difference implementation will be the same as that of
an implementation with exact derivatives.
6.2.2. Other Newton-iterative methods. Other iterative methods may
be used as the solver for the linear equation for the Newton step. Newton-
multigrid [99] is one approach using a stationary iterative method. See also
[6], [111]. If F

is symmetric and positive definite, as it is in the context of
unconstrained minimization, [63], Newton-CG is a possible approach.
When storage is restricted or the problem is very large, GMRES may not
be practical. GMRES(m), for a small m, may not converge rapidly enough
to be useful. CGNR and CGNE offer alternatives, but require evaluation of
transpose-vector products. Bi-CGSTAB and TFQMR should be considered in
all cases where there is not enough storage for GMRES(m) to perform well.
If a transpose is needed, as it will be if CGNR or CGNE is used as the
iterative method, a forward difference formulation is not an option because
approximation of the transpose-vector product by differences is not possible.
Computation of ∇
h
F(x) and its transpose analytically is one option as is
differentiation of F automatically or by hand. The reader may explore this
further in Exercise 6.5.13.
The reader may also be interested in experimenting with the other Krylov
subspace methods described in ¸ 3.6, [78], and [12].
6.3. Newton-GMRES implementation
We provide a MATLAB code nsolgm in the collection that implements
a forward-difference Newton-GMRES algorithm. This algorithm requires
different inputs than nsol. As input we must give the initial iterate, the
function, the vector of termination tolerances as we did for nsol. In addition,
we must provide a method for forming the sequence of forcing terms ¦η
n
¦.
We adjust η as the iteration progresses with a variation of a choice from
[69]. This issue is independent of the particular linear solver and our discussion
is in the general inexact Newton method setting. Setting η to a constant for
the entire iteration is often a reasonable strategy as we see in ¸ 6.4, but the
choice of that constant depends on the problem. If a constant η is too small,
much effort can be wasted in the initial stages of the iteration. The choice in
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 105
[69] is motivated by a desire to avoid such over solving. Over solving means
that the linear equation for the Newton step is solved to a precision far beyond
what is needed to correct the nonlinear iteration. As a measure of the degree
to which the nonlinear iteration approximates the solution we begin with
η
A
n
= γ|F(x
n
)|
2
/|F(x
n−1
)|
2
,
where γ ∈ (0, 1] is a parameter. If η
A
n
is uniformly bounded away from 1,
then setting η
n
= η
A
n
for n > 0 would guarantee q-quadratic convergence by
Theorem 6.1.1. In this way, the most information possible would be extracted
from the inner iteration. In order to specify the choice at n = 0 and bound
the sequence away from 1 we set
η
B
n
=
_
¸
_
¸
_
η
max
, n = 0,
min(η
max
, η
A
n
), n > 0.
(6.18)
In (6.18) the parameter η
max
is an upper limit on the sequence ¦η
n
¦. In [69]
the choices γ = .9 and η
max
= .9999 are used.
It may happen that η
B
n
is small for one or more iterations while x
n
is still
far from the solution. A method of safeguarding was suggested in [69] to avoid
volatile decreases in η
n
. The idea is that if η
n−1
is sufficiently large we do not
let η
n
decrease by much more than a factor of η
n−1
.
η
C
n
=
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
η
max
, n = 0,
min(η
max
, η
A
n
), n > 0, γη
2
n−1
≤ .1,
min(η
max
, max(η
A
n
, γη
2
n−1
)), n > 0, γη
2
n−1
> .1.
(6.19)
The constant .1 is somewhat arbitrary. This safeguarding does improve the
performance of the iteration.
There is a chance that the final iterate will reduce |F| far beyond the
desired level and that the cost of the solution of the linear equation for the last
step will be more accurate than is really needed. This oversolving on the final
step can be controlled comparing the norm of the current nonlinear residual
|F(x
n
)| to the nonlinear residual norm at which the iteration would terminate
τ
t
= τ
a

r
|F(x
0
)|
and bounding η
n
from below by a constant multiple of τ
t
/|F(x
n
)|. The
algorithm nsolgm does this and uses
η
n
= min(η
max
, max(η
C
n
, .5τ
t
/|F(x
n
)|)). (6.20)
Exercise 6.5.9 asks the reader to modify nsolgm so that other choices of
the sequence ¦η
n
¦ can be used.
We use dirder to approximate directional derivatives and use the default
value of h = 10
−7
. In all the examples in this book we use the value γ = .9 as
recommended in [69].
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
106 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Algorithm 6.3.1. nsolgm(x, F, τ, η)
1. r
c
= r
0
= |F(x)|
2
/

N
2. Do while |F(x)|
2
/

N > τ
r
r
0

a
(a) Select η.
(b) fdgmres(s, x, F, η)
(c) x = x +s
(d) Evaluate F(x)
(e) r
+
= |F(x)|
2
/

N, σ = r
+
/r
c
, r
c
= r
+
(f) If |F(x)|
2
≤ τ
r
r
0

a
exit.
Note that the cost of an outer iterate is one evaluation of F to compute
the value at the current iterate and other evaluations of F to compute the
forward differences for each inner iterate. Hence if a large value of η can be
used, the cost of the entire iteration can be greatly reduced. If a maximum
of m GMRES iterations is needed (or GMRES(m) is used as a linear solver)
the storage requirements of Algorithm nsolgm are the m+ 5 vectors x, F(x),
x +hv, F(x +hv), s, and the Krylov basis ¦v
k
¦
m
k=1
.
Since GMRES forms the inner iterates and makes termination decisions
based on scalar products and l
2
norms, we also terminate the outer iteration on
small l
2
norms of the nonlinear residuals. However, because of our applications
to differential and integral equations, we scale the l
2
norm of the nonlinear
residual by a factor of 1/

N so that constant functions will have norms that
are independent of the computational mesh.
For large and poorly conditioned problems, GMRES will have to be
restarted. We illustrate the consequences of this in ¸ 6.4 and ask the reader to
make the necessary modifications to nsolgm in Exercise 6.5.9.
The preconditioners we use in ¸ 6.4 are independent of the outer iteration.
Since the Newton steps for MF(x) = 0 are the same as those for F(x) = 0, it
is equivalent to precondition the nonlinear equation before calling nsolgm.
6.4. Examples for Newton-GMRES
In this section we consider two examples. We revisit the H-equation and solve
a preconditioned nonlinear convection-diffusion equation.
In all the figures we plot the relative nonlinear residual |F(x
n
)|
2
/|F(x
0
)|
2
against the number of function evaluations required by all inner and outer
iterations to compute x
n
. Counts of function evaluations corresponding to
outer iterations are indicated by circles. From these plots one can compare
not only the number of outer iterations, but also the total cost. This enables
us to directly compare the costs of different strategies for selection of η. Note
that if only a single inner iteration is needed, the total cost of the outer iteration
will be two function evaluations since F(x
c
) will be known. One new function
evaluation will be needed to approximate the action of F

(x
c
) on a vector
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 107
and then F(x
+
) will be evaluated to test for termination. We also count
the evaluation of F(x
0
) as an additional cost in the evaluation of x
1
, which
therefore has a minimum cost of three function evaluations. For each example
we compare a constant value of η
n
with the choice given in (6.20) and used as
the default in nsolgm.
6.4.1. Chandrasekhar H-equation. We solve the H-equation on a 100-
point mesh with c = .9 using two schemes for selection of the parameter η.
The initial iterate is the function identically one. We set the parameters in
(6.20) to γ = .9 and η
max
= .25.
We used τ
r
= τ
a
= 10
−6
in this example.
In Fig. 6.1 we plot the progress of the iteration using η = .1 with the solid
line and using the sequence given by (6.20) with the dashed line. We set the
parameters in (6.20) to γ = .9 and η
max
= .25. This choice of η
max
was the
one that did best overall in our experiments on this problem.
We see that the constant η iteration terminates after 12 function evalua-
tions, 4 outer iterations, and roughly 275 thousand floating-point operations.
This is a slightly higher overall cost than the other approach in which ¦η
n
¦ is
given by (6.20), which terminated after 10 function evaluations, 3 outer itera-
tions, and 230 thousand floating-point operations. The chord method with the
different (but very similar for this problem) l

termination condition for the
same problem, reported in ¸ 5.6 incurred a cost of 3.3 million floating-point
operations, slower by a factor of over 1000. This cost is entirely a result of the
computation and factorization of a single Jacobian.
0 2 4 6 8 10 12
10
-8
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 6.1. Newton-GMRES for the H-equation, c = .9.
In Fig. 6.2 the results change for the more ill-conditioned problem with
c = .9999. Here the iteration with η
n
= .1 performed slightly better and
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
108 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
0 5 10 15 20 25
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 6.2. Newton-GMRES for the H-equation, c = .9999
0 2 4 6 8 10 12 14 16 18
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 6.3. Newton-GMRES for the PDE, C = 20.
terminated after 22 function evaluations and 7 outer iterations, and roughly
513 thousand floating-point operations. The iteration with the decreasing ¦η
n
¦
given by (6.20) required 23 function evaluations, 7 outer iterations, and 537
thousand floating-point operations.
6.4.2. Convection-diffusion equation. We consider the partial differen-
tial equation
−∇
2
u +Cu(u
x
+u
y
) = f (6.21)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 109
with homogeneous Dirichlet boundary conditions on the unit square (0, 1)
(0, 1). f has been constructed so that the exact solution was the discretization
of
10xy(1 −x)(1 −y) exp(x
4.5
).
We set C = 20 and u
0
= 0. As in ¸ 3.7 we discretized (6.21) on a 31 31 grid
using centered differences. To be consistent with the second-order accuracy of
the difference scheme we used τ
r
= τ
a
= h
2
, where h = 1/32.
We compare the same possibilities for ¦η
n
¦ as in the previous example and
report the results in Fig. 6.3. In (6.20) we set γ = .9 and η
max
= .5. The
computation with constant η terminated after 19 function evaluations and 4
outer iterates at a cost of roughly 3 million floating-point operations. The
iteration with ¦η
n
¦ given by (6.20) required 16 function evaluations, 4 outer
iterations, and 2.6 million floating-point operations. In the computation we
preconditioned (6.21) with the fast Poisson solver fish2d.
In this case, the choice of ¦η
n
¦ given by (6.20) reduced the number of inner
iterations in the early stages of the iteration and avoided oversolving. In cases
where the initial iterate is very good and only a single Newton iterate might be
needed, however, a large choice of η
max
or constant η may be the more efficient
choice. Examples of such cases include (1) implicit time integration of nonlinear
parabolic partial differential equations or ordinary differential equations in
which the initial iterate is either the converged solution from the previous
time step or the output of a predictor and (2) solution of problems in which
the initial iterate is an interpolant of the solution from a coarser mesh.
Preconditioning is crucial for good performance. When preconditioning
was not done, the iteration for C = 20 required more than 80 function
evaluations to converge. In Exercise 6.5.9 the reader is asked to explore this.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
110 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
6.5. Exercises on inexact Newton methods
6.5.1. Verify (6.8).
6.5.2. Prove Proposition 6.1.1 by showing that if the standard assumptions
hold, 0 < < 1, and x is sufficiently near x

then
|F(x)|/(1 +) ≤ |F

(x

)e| ≤ (1 +)|F(x)|.
6.5.3. Verify (6.14).
6.5.4. Prove Theorems 6.2.2 and 6.2.1.
6.5.5. Prove Theorems 6.1.5 and 6.1.6.
6.5.6. Can anything like Proposition 6.2.1 be proved for a finite-difference Bi-
CGSTAB or TFQMR linear solver? If not, what are some possible
implications?
6.5.7. Give an example in which F(x
n
) → 0 q-linearly but x
n
does not converge
to x

q-linearly.
6.5.8. In the DASPK code for integration of differential algebraic equations,
[23], preconditioned Newton-GMRES is implemented in such a way
that the data needed for the preconditioner is only computed when
needed, which may be less often that with each outer iteration. Discuss
this strategy and when it might be useful. Is there a relation to the
Shamanskii method?
6.5.9. Duplicate the results in ¸ 6.4. For the H-equation experiment with
various values of c, such as c = .5, .99999, .999999, and for the convection-
diffusion equation various values of C such as C = 1, 10, 40 and how
preconditioning affects the iteration. Experiment with the sequence of
forcing terms ¦η
n
¦. How do the choices η
n
= 1/(n + 1), η
n
= 10
−4
[32], η
n
= 2
1−n
[24], η
n
= min(|F(x
n
)|
2
, (n + 2)
−1
) [56], η
n
= .05, and
η
n
= .75 affect the results? Present the results of your experiments as
graphical iteration histories.
6.5.10. For the H-equation example in ¸ 6.4, vary the parameter c and see how
the performance of Newton-GMRES with the various choices of ¦η
n
¦ is
affected. What happens when c = 1? See [119] for an explanation of
this.
6.5.11. Are the converged solutions for the preconditioned and unpreconditioned
convection-diffusion example in ¸ 6.4 equally accurate?
6.5.12. For the convection-diffusion example in ¸ 6.4, how is the performance
affected if GMRES(m) is used instead of GMRES?
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INEXACT NEWTON METHODS 111
6.5.13. Apply Newton-CGNR, Newton-Bi-CGSTAB, or Newton-TFQMR to the
two examples in ¸ 6.4. How will you deal with the need for a transpose in
the case of CGNR? How does the performance (in terms of time, floating-
point operations, function evaluations) compare to Newton-GMRES?
Experiment with the forcing terms. The MATLAB codes fdkrylov,
which allows you to choose from a variety of forward difference Krylov
methods, and nsola, which we describe more fully in Chapter 8, might
be of use in this exercise.
6.5.14. If you have done Exercise 5.7.25, apply nsolgm to the same two-point
boundary value problem and compare the performance.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
112 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 7
Broyden’s method
Quasi-Newton methods maintain approximations of the solution x

and the
Jacobian at the solution F

(x

) as the iteration progresses. If x
c
and B
c
are
the current approximate solution and Jacobian, then
x
+
= x
c
−B
−1
c
F(x
c
). (7.1)
After the computation of x
+
, B
c
is updated to form B
+
. The construction of
B
+
determines the quasi-Newton method.
The advantages of such methods, as we shall see in ¸ 7.3, is that solution of
equations with the quasi-Newton approximation to the Jacobian is often much
cheaper than using F

(x
n
) as the coefficient matrix. In fact, if the Jacobian is
dense, the cost is little more than that of the chord method. The method we
present in this chapter, Broyden’s method, is locally superlinearly convergent,
and hence is a very powerful alternative to Newton’s method or the chord
method.
In comparison with Newton-iterative methods, quasi-Newton methods
require only one function evaluation for each nonlinear iterate; there is no cost
in function evaluations associated with an inner iteration. Hence, if a good
preconditioner (initial approximation to F

(x

)) can be found, these methods
could have an advantage in terms of function evaluation cost over Newton-
iterative methods.
Broyden’s method [26] computes B
+
by
B
+
= B
c
+
(y −B
c
s)s
T
s
T
s
= B
c
+
F(x
+
)s
T
s
T
s
. (7.2)
In (7.2) y = F(x
+
) −F(x
c
) and s = x
+
−x
c
.
Broyden’s method is an example of a secant update. This means that the
updated approximation to F

(x

) satisfies the secant equation
B
+
s = y. (7.3)
In one dimension, (7.3) uniquely specifies the classical secant method. For
equations in several variables (7.3) alone is a system of N equations in N
2
113
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
114 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
unknowns. The role of the secant equation is very deep. See [62] and [64] for
discussion of this topic. In this book we focus on Broyden’s method and our
methods of analysis are not as general as those in [62] and [64]. In this chapter
we will assume that the function values are accurate and refer the reader to [66],
[65], and [114] for discussions of quasi-Newton methods with inaccurate data.
Locally superlinearly convergent secant methods can be designed to maintain
the sparsity pattern, symmetry, or positive definiteness, of the approximate
Jacobian, and we refer the reader to [28], [63], [62], and [64] for discussion and
analysis of several such methods. More subtle structural properties can also
be preserved by secant methods. Some examples can be found in [101], [96],
[95], [113], and [116].
Broyden’s method is also applicable to linear equations [29], [82], [84],
[104], [143], [130], Ax = b, where B ≈ A is updated. One can also apply the
method to linear least squares problems [84], [104]. The analysis is somewhat
clearer in the linear case and the plan of this chapter is to present results for
the linear case for both the theory and the examples. One interesting result,
which we will not discuss further, is that for linear problems Broyden’s method
converges in 2N iterations [29], [82], [84], [143].
We will express our results in terms of the error in the Jacobian
E = B −F

(x

) (7.4)
and the step s = x
+
−x
c
. While we could use E = B−F

(x), (7.4) is consistent
with our use of e = x − x

. When indexing of the steps is necessary, we will
write
s
n
= x
n+1
−x
n
.
In this chapter we show that if the data x
0
and B
0
are sufficiently good,
the Broyden iteration will converge q-superlinearly to the root.
7.1. The Dennis–Mor´e condition
In this section we consider general iterative methods of the form
x
n+1
= x
n
−B
−1
n
F(x
n
) (7.5)
where B
n
= F

(x

)+E
n
≈ F

(x

) is generated by some method (not necessarily
Broyden’s method).
Verification of q-superlinear convergence cannot be done by the direct
approaches used in Chapters 5 and 6. We must relate the superlinear
convergence condition that η
n
→ 0 in Theorem 6.1.2 to a more subtle, but
easier to verify condition in order to analyze Broyden’s method. This technique
is based on verification of the Dennis–Mor´e condition [61], [60] on the sequence
of steps ¦s
n
¦ and errors in the Jacobian ¦E
n
¦
lim
n→∞
|E
n
s
n
|
|s
n
|
= 0. (7.6)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 115
The main result in this section, Theorem 7.1.1 and its corollary for linear
problems, are used to prove superlinear convergence for Broyden’s method in
this chapter and many other quasi-Newton methods. See [61], [28], [62], and
[64] for several more examples. Our formulation of the Dennis–Mor´e result is
a bit less general than that in [60] or [63].
Theorem 7.1.1. Let the standard assumptions hold, let ¦B
n
¦ be a sequence
of nonsingular N N matrices, let x
0
∈ R
N
be given and let ¦x
n
¦

n=1
be given
by (7.5). Assume that x
n
,= x

for any n. Then x
n
→ x

q-superlinearly if
and only if x
n
→ x

and the Dennis–Mor´e condition (7.6) holds.
Proof. Since
−F(x
n
) = B
n
s
n
= F

(x

)s
n
+E
n
s
n
we have
E
n
s
n
= −F

(x

)s
n
−F(x
n
) = −F

(x

)e
n+1
+F

(x

)e
n
−F(x
n
). (7.7)
We use the fundamental theorem of calculus and the standard assumptions to
obtain
F

(x

)e
n
−F(x
n
) =
_
1
0
(F

(x

) −F

(x

+te
n
))e
n
dt
and hence
|F

(x

)e
n
−F(x
n
)| ≤ γ|e
n
|
2
/2.
Therefore, by (7.7)
|E
n
s
n
| ≤ |F

(x

)e
n+1
| +γ|e
n
|
2
/2. (7.8)
Now, if x
n
→ x

q-superlinearly, then for n sufficiently large
|s
n
|/2 ≤ |e
n
| ≤ 2|s
n
|. (7.9)
The assumption that x
n
,= x

for any n implies that the sequence
ν
n
= |e
n+1
|/|e
n
| (7.10)
is defined and and ν
n
→ 0 by superlinear convergence of ¦x
n
¦. By (7.9) we
have
|e
n+1
| = ν
n
|e
n
| ≤ 2ν
n
|s
n
|,
and hence (7.8) implies that
|E
n
s
n
| ≤ (2|F

(x

)|ν
n
+γ|e
n
|)|s
n
|
which implies the Dennis-Mor´e condition (7.6).
Conversely, assume that x
n
→ x

and that (7.6) holds. Let
µ
n
=
|E
n
s
n
|
|s
n
|
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
116 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
We have, since
E
n
s
n
= (B
n
−F

(x

))s
n
= −F(x
n
) −F

(x

)s
n
, (7.11)
|s
n
||F

(x

)
−1
|
−1
≤ |F

(x

)s
n
| ≤ |E
n
s
n
| +|F(x
n
)|
= µ
n
|s
n
| +|F(x
n
)|.
(7.12)
Since µ
n
→ 0 by assumption,
µ
n
≤ |F

(x

)
−1
|
−1
/2, (7.13)
for n sufficiently large. We assume now that n is large enough so that (7.13)
holds. Hence,
|s
n
| ≤ 2|F

(x

)
−1
||F(x
n
)|. (7.14)
Moreover, using the standard assumptions and (7.11) again,
|F(x
n
) +F

(x
n
)s
n
| ≤ |F(x
n
) +F

(x

)s
n
| +|(F

(x

) −F

(x
n
))s
n
|
≤ (µ
n
+γ|e
n
|)|s
n
|.
(7.15)
Since x
n
→ x

by assumption, η
n
→ 0 where
η
n
= 2|F

(x

)
−1
|(µ
n
+γ|e
n
|).
Combining (7.14) and (7.15) implies that
|F(x
n
) +F

(x
n
)s
n
| ≤ η
n
|F(x
n
)|, (7.16)
which is the inexact Newton condition. Since η
n
→ 0, the proof is complete
by Theorem 6.1.2.
7.2. Convergence analysis
Our convergence analysis is based on an approach to infinite-dimensional
problems from [97]. That paper was the basis for [168], [104], and [115], and
we employ notation from all four of these references. We will use the l
2
norm
throughout our discussion. We begin with a lemma from [104].
Lemma 7.2.1. Let 0 <
ˆ
θ < 1 and let
¦θ
n
¦

n=0
⊂ (
ˆ
θ, 2 −
ˆ
θ).
Let ¦
n
¦

n=0
⊂ R
N
be such that

n
|
n
|
2
< ∞,
and let ¦η
n
¦

n=0
be a set of vectors in R
N
such that |η
n
|
2
is either 1 or 0 for
all n. Let ψ
0
∈ R
N
be given. If ¦ψ
n
¦

n=1
is given by
ψ
n+1
= ψ
n
−θ
n

T
n
ψ
n

n
+
n
(7.17)
then
lim
n→∞
η
T
n
ψ
n
= 0. (7.18)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 117
Proof. We first consider the case in which
n
= 0 for all n. In that case, we
can use the fact that
θ
n
(2 −θ
n
) >
ˆ
θ
2
> 0
to show that the sequence ¦ψ
n
¦ is bounded in l
2
-norm by |ψ
0
|
2
and, in fact,

n+1
|
2
2
≤ |ψ
n
|
2
2
−θ
n
(2 −θ
n
)(η
T
n
ψ
n
)
2
≤ |ψ
n
|
2
2

ˆ
θ
2

T
n
ψ
n
)
2
≤ |ψ
n
|
2
2
.
Therefore, for any M > 0,
M

n=0

T
n
ψ
n
)
2


0
|
2
2
−|ψ
M+1
|
2
2
ˆ
θ
2


0
|
2
ˆ
θ
2
.
We let M → ∞ to obtain

n=0

T
n
ψ
n
)
2
< ∞. (7.19)
Convergence of the series in (7.19) implies convergence to zero of the terms in
the series, which is (7.18).
To prove the result for
n
,= 0 we use the inequality
_
a
2
−b
2
≤ a −
b
2
2a
, (7.20)
which is valid for a > 0 and [b[ ≤ a. This inequality is used often in the analysis
of quasi-Newton methods [63]. From (7.20) we conclude that if ψ
n
,= 0 then

n
−θ
n

T
n
ψ
n

n
|
2

_

n
|
2
2
−θ
n
(2 −θ
n
)(η
T
n
ψ
n
)
2
≤ |ψ
n
|
2

θ
n
(2 −θ
n
)(η
T
n
ψ
n
)
2
2|ψ
n
|
2
.
Hence if ψ
n
,= 0

n+1
|
2
≤ |ψ
n
|
2

θ
n
(2 −θ
n
)(η
T
n
ψ
n
)
2
2|ψ
n
|
2
+|
n
|
2
. (7.21)
Hence

T
n
ψ
n
)
2

2|ψ
n
|
2
θ
n
(2 −θ
n
)
(|ψ
n
|
2
−|ψ
n+1
|
2
+|
n
|
2
), (7.22)
which holds even if ψ
n
= 0.
From (7.21) we conclude that

n+1
|
2
≤ µ,
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
118 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
where
µ =

i=0
|
i
|
2
+|ψ
0
|
2
.
Hence

M
n=0

T
n
ψ
n
)
2


ˆ
θ
2
M

n=0
(|ψ
n
|
2
−|ψ
n+1
|
2
+|
n
|
2
)
=

ˆ
θ
2
_

0
|
2
−|ψ
M+1
|
2
+
M

n=0
|
n
|
2
_


2
ˆ
θ
2
.
Hence (7.19) holds and the proof is complete.
7.2.1. Linear problems. We may use Lemma 7.2.1 to prove a result from
[130] on the convergence of Broyden’s method for linear problems. The role of
the parameter
ˆ
θ and the sequence ¦θ
n
¦ in Lemma 7.2.1 is nicely illustrated by
this application of the lemma.
In this section F(x) = Ax−b and B
n
= A+E
n
. The standard assumptions
in this case reduce to nonsingularity of A. Our first task is to show that
the errors in the Broyden approximations to A do not grow. This property,
called bounded deterioration [28], is an important part of the convergence
analysis. The reader should compare the statement of the result with that
in the nonlinear case.
In the linear case, the Broyden update has a very simple form. We will
consider a modified update
B
+
= B
c

c
(y −B
c
s)s
T
s
T
s
= B
c

c
(Ax
+
−b)s
T
s
T
s
. (7.23)
If x
c
and B
c
are the current approximations to x

= A
−1
b and A and
s = x
+
−x
c
then
y −B
c
s = (Ax
+
−Ax
c
) −B
c
(x
+
−x
c
) = −E
c
s
and therefore
B
+
= B
c

c
(y −B
c
s)s
T
|s|
2
2
= B
c
−θ
c
(E
c
s)s
T
|s|
2
2
. (7.24)
Lemma 7.2.2. Let A be nonsingular, θ
c
∈ [0, 2], and x
c
∈ R
N
be given. Let
B
c
be nonsingular and let B
+
be formed by the Broyden update (7.23). Then
|E
+
|
2
≤ |E
c
|
2
. (7.25)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 119
Proof. By subtracting A from both sides of (7.24) we have
E
+
= E
c
−θ
c
(E
c
s)s
T
|s|
2
2
= E
c
(I −θ
c
P
s
),
(7.26)
where P
s
is the orthogonal projector
P
s
=
ss
T
|s|
2
2
. (7.27)
This completes the proof as
|E
+
|
2
≤ |E
c
|
2
|I −θ
c
P
s
|
2
and |I −θ
c
P
s
|
2
≤ 1 by orthogonality of P
s
and the fact that 0 ≤ θ
c
≤ 2.
Now, θ
c
can always be selected to make B
+
nonsingular. In [130] one
suggestion was
θ
c
=
_
_
_
1, [γ
c
[ ≥ σ,
1 −sign(γ
c

1 −γ
c
, otherwise,
where
γ
c
=
(B
−1
c
y)
T
s
|s|
2
2
=
(B
−1
c
As)
T
s
|s|
2
2
and σ ∈ (0, 1) is fixed. However the results in [130] assume only that the
sequence ¦θ
n
¦ satisfies the hypotheses of Lemma 7.2.1 for some
ˆ
θ ∈ (0, 1) and
that θ
c
is always chosen so that B
+
is nonsingular.
For linear problems one can show that the Dennis–Mor´e condition implies
convergence and hence the assumption that convergence takes place is not
needed. Specifically
Proposition 7.2.1. Let A be nonsingular. Assume that ¦θ
n
¦ satisfies
the hypotheses of Lemma 7.2.1 for some
ˆ
θ ∈ (0, 1). If ¦θ
n
¦ is such that
the matrices B
n
obtained by (7.23) are nonsingular and ¦x
n
¦ is given by the
modified Broyden iteration
x
n+1
= x
n
−B
−1
n
(Ax
n
−b) (7.28)
then the Dennis–Mor´e condition implies convergence of ¦x
n
¦ to x

= A
−1
b.
We leave the proof as Exercise 7.5.1.
The superlinear convergence result is remarkable in that the initial iterate
need not be near the root.
Theorem 7.2.1. Let A be nonsingular. Assume that ¦θ
n
¦ satisfies the
hypotheses of Lemma 7.2.1 for some
ˆ
θ ∈ (0, 1). If ¦θ
n
¦ is such the matrices B
n
obtained by (7.23) are nonsingular then the modified Broyden iteration (7.28)
converges q-superlinearly to x

= A
−1
b for every initial iterate x
0
.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
120 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Proof. Our approach is to show that for every φ ∈ R
N
that
lim
n→∞
φ
T
_
E
n
s
n
|s
n
|
2
_
= 0. (7.29)
Assuming that (7.29) holds, setting φ = e
j
, the unit vector in the jth
coordinate direction, shows that the j component of
E
n
s
n
|s
n
|
2
has limit zero.
As j was arbitrary, we have
lim
n→∞
E
n
s
n
|s
n
|
2
= 0
which is the Dennis–Mor´e condition. This will imply q-superlinear convergence
by Theorem 7.1.1 and Proposition 7.2.1.
Let φ ∈ R
N
be given. Let
P
n
=
s
n
s
T
n
|s
n
|
2
2
.
Since for all v ∈ R
N
and all n
v
T
(E
T
n
φ) = φ
T
(E
n
v)
and, by (7.26)
E
T
n+1
φ = (I −θ
n
P
n
)E
T
n
φ,
we may invoke Lemma 7.2.1 with
η
n
= s
n
/|s
n
|
2
, ψ
n
= E
T
n
φ, and
n
= 0
to conclude that
η
T
n
ψ
n
=
(E
T
n
φ)
T
s
n
|s
n
|
2
= φ
T
E
n
s
n
|s
n
|
2
→ 0. (7.30)
This completes the proof.
The result here is not a local convergence result. It is global in the sense
that q-superlinear convergence takes place independently of the initial choices
of x and B. It is known [29], [82], [84], [143], that the Broyden iteration will
terminate in 2N steps in the linear case.
7.2.2. Nonlinear problems. Our analysis of the nonlinear case is different
from the classical work [28], [63]. As indicated in the preface, we provide a
proof based on ideas from [97], [115], and [104], that does not use the Frobenius
norm and extends to various infinite-dimensional settings.
For nonlinear problems our analysis is local and we set θ
n
= 1. However
we must show that the sequence ¦B
n
¦ of Broyden updates exists and remains
nonsingular. We make a definition that allows us to compactly express this
fact.
Definition 7.2.1. We say that the Broyden sequence (¦x
n
¦, ¦B
n
¦) for the
data (F, x
0
, B
0
) exists if F is defined at x
n
for all n, B
n
is nonsingular for all
n, and x
n+1
and B
n+1
can be computed from x
n
and B
n
using (7.1) and (7.2).
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 121
We will show that if the standard assumptions hold and x
0
and B
0
are
sufficiently good approximations to x

and F

(x

) then the Broyden sequence
exists for the data (F, x
0
, B
0
) and that the Broyden iterates converge q-
superlinearly to x

. The development of the proof is complicated. We first
prove a bounded deterioration result like Lemma 7.2.2. However in the
nonlinear case, the errors in the Jacobian approximation can grow as the
iteration progresses. We then show that if the initial data is sufficiently good,
this growth can be limited and therefore the Broyden sequence exists and
the Broyden iterates converges to x

at least q-linearly. Finally we verify the
Dennis–Mor´e condition (7.6), which, together with the q-linear convergence
proved earlier completes the proof of local q-superlinear convergence.
Bounded deterioration. Our first task is to show that bounded deteriora-
tion holds. Unlike the linear case, the sequence ¦|E
n
|¦ need not be monotone
non-increasing. However, the possible increase can be bounded in terms of the
errors in the current and new iterates.
Theorem 7.2.2. Let the standard assumptions hold. Let x
c
∈ Ω and a
nonsingular matrix B
c
be given. Assume that
x
+
= x
c
−B
−1
c
F(x
c
) = x
c
+s ∈ Ω
and B
+
is given by (7.2). Then
|E
+
|
2
≤ |E
c
| +γ(|e
c
|
2
+|e
+
|
2
)/2 (7.31)
Proof. Note that (7.7) implies that
E
c
s = −F(x
+
) + (F(x
+
) −F(x
c
) −F

(x

)s)
= −F(x
+
) +
_
1
0
(F

(x
c
+ts) −F

(x

))s dt.
(7.32)
We begin by writing (7.32) as
F(x
+
) = −E
c
s +
_
1
0
(F

(x
c
+ts) −F

(x

))s dt
and then using (7.2) to obtain
E
+
= E
c
(I −P
s
) +
(∆
c
s)s
T
|s|
2
2
where P
s
is is given by (7.27) and

c
=
_
1
0
(F

(x
c
+ts) −F

(x

)) dt.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
122 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Hence, by the standard assumptions
|∆
c
|
2
≤ γ(|e
c
|
2
+|e
+
|
2
)/2.
Just as in the proof of Lemma 7.2.2 for the linear case
|E
+
|
2
≤ |E
c
|
2
+|∆
c
|
2
and the proof is complete.
Local q-linear convergence. Theorem 7.2.3 states that Broyden’s method
is locally q-linearly convergent. The proof is more complex than similar results
in Chapters 5 and 6. We must explore the relation between the size of |E
0
|
needed to enforce q-linear convergence and the q-factor.
Theorem 7.2.3. Let the standard assumptions hold. Let r ∈ (0, 1) be
given. Then there are δ and δ
B
such that if x
0
∈ B(δ) and |E
0
|
2
< δ
B
the
Broyden sequence for the data (F, x
0
, B
0
) exists and x
n
→ x

q-linearly with
q-factor at most r.
Proof. Let δ and δ
1
be small enough so that the conclusions of Theo-
rem 5.4.3 hold. Then, reducing δ and δ
1
further if needed, the q-factor is at
most
K
A
(δ +δ
1
) ≤ r,
where K
A
is from the statement of Theorem 5.4.3.
We will prove the result by choosing δ
B
and reducing δ if necessary so that
|E
0
|
2
≤ δ
B
will imply that |E
n
|
2
≤ δ
1
for all n, which will prove the result.
To do this reduce δ if needed so that
δ
2
=
γ(1 +r)δ
2(1 −r)
< δ
1
(7.33)
and set
δ
B
= δ
1
−δ
2
. (7.34)
We show that |E
n
|
2
≤ δ
1
by induction. Since |E
0
|
2
< δ
B
< δ
1
we may
begin the induction. Now assume that |E
k
|
2
< δ
1
for all 0 ≤ k ≤ n. By
Theorem 7.2.2,
|E
n+1
|
2
≤ |E
n
|
2
+γ(|e
n+1
|
2
+|e
n
|
2
)/2.
Since |E
n
|
2
< δ
1
, |e
n+1
|
2
≤ r|e
n
|
2
and therefore
|E
n+1
|
2
≤ |E
n
|
2
+γ(1 +r)|e
n
|
2
/2 ≤ |E
n
|
2
+γ(1 +r)r
n
δ/2.
We can estimate |E
n
|
2
, |E
n−1
|
2
, . . . in the same way to obtain
|E
n+1
|
2
≤ |E
0
|
2
+
(1 +r)γδ
2
n

j=0
r
j
≤ δ
B
+
(1 +r)γδ
2(1 −r)
≤ δ
1
,
which completes the proof.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 123
Verification of the Dennis–Mor´e condition. The final task is to verify
the Dennis–Mor´e condition.
Theorem 7.2.4. Let the standard assumptions hold. Then there are δ and
δ
B
such that if x
0
∈ B(δ) and |E
0
|
2
< δ
B
the Broyden sequence for the data
(F, x
0
, B
0
) exists and x
n
→ x

q-superlinearly.
Proof. Let δ and δ
B
be such that the conclusions of Theorem 7.2.3 hold.
By Theorem 7.1.1 we need only show that Dennis–Mor´e condition (7.6) holds
to complete the proof.
As in the proof of Theorem 7.2.1 let
P
n
=
s
n
s
T
n
|s
n
|
2
2
.
Set

n
=
_
1
0
(F

(x
n
+ts
n
) −F

(x

)) dt.
Let φ ∈ R
N
be arbitrary. Note that
E
T
n+1
φ = (I −P
n
)E
T
n
φ +P
n

T
n
φ.
We wish to apply Lemma 7.2.1 with
ψ
n
= E
T
n
φ, η
n
= s
n
/|s
n
|
2
, and
n
= P
n

T
n
φ.
The hypothesis in the lemma that

n
|
n
|
2
< ∞
holds since Theorem 7.2.3 and the standard assumptions imply
|∆
n
|
2
≤ γ(1 +r)r
n
δ/2.
Hence Lemma 7.2.1 implies that
η
T
n
ψ
n
=
(E
T
n
φ)
T
s
n
|s
n
|
2
= φ
T
E
n
s
n
|s
n
|
2
→ 0.
This, as in the proof of Theorem 7.2.1, implies the Dennis–Mor´e condition and
completes the proof.
7.3. Implementation of Broyden’s method
The implementation of Broyden’s method that we advocate here is related
to one that has been used in computational fluid mechanics [73]. Some
such methods are called limited memory formulations in the optimization
literature [138], [142], [31]. In these methods, after the storage available for
the iteration is exhausted, the oldest of the stored vectors is replaced by the
most recent. Another approach, restarting, clears the storage and starts over.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
124 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
This latter approach is similar to the restarted GMRES iteration. Our basic
implementation is a nonlinear version of the implementation for linear problems
in [67]. We use a restarted approach, though a limited memory approach could
be used as well.
We begin by showing how the initial approximation to F

(x

), B
0
, may be
regarded as a preconditioner and incorporated into the defintion of F just as
preconditioners were in the implementation of Newton-GMRES in ¸ 6.4.
Lemma 7.3.1. Assume that the Broyden sequence (¦x
n
¦, ¦B
n
¦) for the data
(F, x
0
, B
0
) exists. Then the Broyden sequence (¦y
n
¦, ¦C
n
¦) exists for the data
(B
−1
0
F, x
0
, I) and
x
n
= y
n
, B
n
= B
0
C
n
(7.35)
for all n.
Proof. We proceed by induction. Equation (7.35) holds by definition when
n = 0. If (7.35) holds for a given n, then
y
n+1
= y
n
−C
−1
n
B
−1
0
F(y
n
)
= x
n
−(B
0
C
n
)
−1
F(x
n
) = x
n
−B
−1
n
F(x
n
)
= x
n+1
.
(7.36)
Now s
n
= x
n+1
−x
n
= y
n+1
−y
n
by (7.36). Hence
B
n+1
= B
n
+
F(x
n+1
)s
T
n
|s
n
|
2
2
= B
0
C
n
+B
0
B
−1
0
F(y
n+1
)s
T
n
|s
n
|
2
2
= B
0
C
n+1
.
This completes the proof.
The consequence of this proof for implementation is that we can assume
that B
0
= I. If a better choice for B
0
exists, we can incorporate that into the
function evaluation. Our plan will be to store B
−1
n
in a compact and easy to
apply form. The basis for this is the following simple formula [68], [176], [177],
[11].
Proposition 7.3.1. Let B be a nonsingular N N matrix and let
u, v ∈ R
N
. Then B + uv
T
is invertible if and only if 1 + v
T
B
−1
u ,= 0. In
this case,
(B +uv
T
)
−1
=
_
I −
(B
−1
u)v
T
1 +v
T
B
−1
u
_
B
−1
. (7.37)
The expression (7.37) is called the Sherman–Morrison formula. We leave
the proof to Exercise 7.5.2.
In the context of a sequence of Broyden updates ¦B
n
¦ we have for n ≥ 0,
B
n+1
= B
n
+u
n
v
T
n
,
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 125
where
u
n
= F(x
n+1
)/|s
n
|
2
and v
n
= s
n
/|s
n
|
2
.
Setting
w
n
= (B
−1
n
u
n
)/(1 +v
T
n
B
−1
n
u
n
)
we see that, if B
0
= I,
B
−1
n
= (I −w
n−1
v
T
n−1
)(I −w
n−2
v
T
n−2
) . . . (I −w
0
v
T
0
)
=

n−1
j=0
(I −w
j
v
T
j
).
(7.38)
Since the empty matrix product is the identity, (7.38) is valid for n ≥ 0.
Hence the action of B
−1
n
on F(x
n
) (i.e., the computation of the Broyden
step) can be computed from the 2n vectors ¦w
j
, v
j
¦
n−1
j=0
at a cost of O(Nn)
floating-point operations. Moreover, the Broyden step for the following
iteration is
s
n
= −B
−1
n
F(x
n
) = −

n−1
j=0
(I −w
j
v
T
j
)F(x
n
). (7.39)
Since the product

n−2
j=0
(I −w
j
v
T
j
)F(x
n
)
must also be computed as part of the computation of w
n−1
we can combine
the computation of w
n−1
and s
n
as follows:
w =

n−2
j=0
(I −w
j
v
T
j
)F(x
n
)
w
n−1
= C
w
w where C
w
= (|s
n−1
|
2
+v
T
n−1
w)
−1
s
n
= −(I −w
n−1
v
T
n−1
)w.
(7.40)
One may carry this one important step further [67] and eliminate the need
to store the sequence ¦w
n
¦. Note that (7.40) implies that for n ≥ 1
s
n
= −w +C
w
w(v
T
n−1
w) = w(−1 +C
w
(C
−1
w
−|s
n−1
|
2
))
= −C
w
w|s
n−1
|
2
= −|s
n−1
|
2
w
n−1
.
Hence, for n ≥ 0,
w
n
= −s
n+1
/|s
n
|
2
. (7.41)
Therefore, one need only store the steps ¦s
n
¦ and their norms to construct
the sequence ¦w
n
¦. In fact, we can write (7.38) as
B
−1
n
=

n−1
j=0
_
I +
s
j+1
s
T
j
|s
j
|
2
2
_
. (7.42)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
126 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
We cannot use (7.42) and (7.39) directly to compute s
n+1
because s
n+1
appears
on both sides of the equation
s
n+1
= −
_
I +
s
n+1
s
T
n
|s
n
|
2
2
_

n−1
j=0
_
I +
s
j+1
s
T
j
|s
j
|
2
2
_
F(x
n+1
)
= −
_
I +
s
n+1
s
T
n
|s
n
|
2
2
_
B
−1
n
F(x
n+1
).
(7.43)
Instead, we solve (7.43) for s
n+1
to obtain
s
n+1
= −
B
−1
n
F(x
n+1
)
1 +s
T
n
B
−1
n
F(x
n+1
)/|s
n
|
2
2
. (7.44)
By Proposition 7.3.1, the denominator in (7.44)
1 +s
T
n
B
−1
n
F(x
n+1
)/|s
n
|
2
2
= 1 +v
T
n
B
−1
n
u
n
is nonzero unless B
n+1
is singular.
The storage requirement, of n +O(1) vectors for the nth iterate, is of the
same order as GMRES, making Broyden’s method competitive in storage with
GMRES as a linear solver [67]. Our implementation is similar to GMRES(m),
restarting with B
0
when storage is exhausted.
We can use (7.42) and (7.44) to describe an algorithm. The termination
criteria are the same as for Algorithm nsol in ¸ 5.6. We use a restarting
approach in the algorithm, adding to the input list a limit nmax on the number
of Broyden iterations taken before a restart and a limit maxit on the number
of nonlinear iterations. Note that B
0
= I is implicit in the algorithm. Our
looping convention is that the loop in step 2(e)ii is skipped if n = 0; this is
consistent with setting the empty matrix product to the identity. Our notation
and algorithmic framework are based on [67].
Algorithm 7.3.1. brsol(x, F, τ, maxit, nmax)
1. r
0
= |F(x)|
2
, n = −1,
s
0
= −F(x), itc = 0.
2. Do while itc < maxit
(a) n = n + 1; itc = itc + 1
(b) x = x +s
n
(c) Evaluate F(x)
(d) If |F(x)|
2
≤ τ
r
r
0

a
exit.
(e) if n < nmax then
i. z = −F(x)
ii. for j = 0, n −1
z = z +s
j+1
s
T
j
z/|s
j
|
2
2
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 127
iii. s
n+1
= z/(1 −s
T
n
z/|s
n
|
2
2
)
(f) if n = nmax then
n = −1; s = −F(x);
If n < nmax, then the nth iteration of brsol, which computes x
n
and
s
n+1
, requires O(nN) floating-point operations and storage of n + 3 vectors,
the steps ¦s
j
¦
n
j=0
, x, z, and F(x), where z and F(x) may occupy the same
storage location. This requirement of one more vector than the algorithm
in [67] needed is a result of the nonlinearity. If n = nmax, the iteration
will restart with x
nmax
, not compute s
nmax+1
, and therefore need storage for
nmax+2 vectors. Our implementation of brsol in the collection of MATLAB
codes stores z and F separately in the interest of clarity and therefore requires
nmax + 3 vectors of storage.
The Sherman–Morrison approach in Algorithm brsol is more efficient, in
terms of both time and storage, than the dense matrix approach proposed
in [85]. However the dense matrix approach makes it possible to detect
ill conditioning in the approximate Jacobians. If the data is sufficiently
good, bounded deterioration implies that the Broyden matrices will be well
conditioned and superlinear convergence implies that only a few iterates will
be needed. In view of all this, we feel that the approach of Algorithm brsol
is the best approach, especially for large problems.
A MATLAB implementation of brsol is provided in the collection of
MATLAB codes.
7.4. Examples for Broyden’s method
In all the examples in this section we use the l
2
norm multiplied by 1/

N.
We use the MATLAB code brsol.
7.4.1. Linear problems. The theory in this chapter indicates that Broy-
den’s method can be viewed as an alternative to GMRES as an iterative method
for nonsymmetric linear equations. This point has been explored in some detail
in [67], where more elaborate numerical experiments that those given here are
presented, and in several papers on infinite-dimensional problems as well [91],
[104], [120]. As an example we consider the linear PDE (3.33) from ¸ 3.7. We
use the same mesh, right-hand side, coefficients, and solution as the example
in ¸ 3.7.
We compare Broyden’s method without restarts (solid line) to Broyden’s
method with restarts every three iterates (dashed line). For linear problems,
we allow for increases in the nonlinear residual. We set τ
a
= τ
r
= h
2
where
h = 1/32. In Fig. 7.1 we present results for (3.33) preconditioned with the
Poisson solver fish2d.
Broyden’s method terminated after 9 iterations at a cost of roughly 1.5
million floating point operations, a slightly higher cost than the GMRES
solution. The restarted iteration required 24 iterations and 3.5 million floating
point operations. One can also see highly irregular performance from the
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
128 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
restarted method.
In the linear case [67] one can store only the residuals and recover the
terminal iterate upon exit. Storage of the residual, the vector z, and the
sequence of steps saves a vector over the nonlinear case.
0 5 10 15 20 25
10
-3
10
-2
10
-1
10
0
10
1
Iterations
R
e
l
a
t
i
v
e

R
e
s
i
d
u
a
l
Fig. 7.1. Broyden’s method for the linear PDE.
7.4.2. Nonlinear problems. We consider the H-equation from Chapters 5
and 6 and a nonlinear elliptic partial differential equation. Broyden’s method
and Newton-GMRES both require more storage as iterations progress, Newton-
GMRES as the inner iteration progresses and Broyden’s method as the outer
iteration progresses. The cost in function evaluations for Broyden’s method is
one for each nonlinear iteration and for Newton-GMRES one for each outer
iteration and one for each inner iteration. Hence, if function evaluations are
very expensive, the cost of a solution by Broyden’s method could be much less
than that of one by Newton-GMRES. If storage is at a premium, however,
and the number of nonlinear iterations is large, Broyden’s method could be at
a disadvantage as it might need to be restarted many times. The examples
in this section, as well as those in ¸ 8.4 illustrate the differences in the two
methods. Our general recommendation is to try both.
Chandrasekhar H-equation. As before we solve the H-equation on a 100-
point mesh with c = .9 (Fig. 7.2) and c = .9999 (Fig. 7.3). We compare
Broyden’s method without restarts (solid line) to Broyden’s method with
restarts every three iterates (dashed line). For c = .9 both methods required six
iterations for convergence. The implementation without restarts required 153
thousand floating-point operations and the restarted method 150, a substantial
improvement over the Newton-GMRES implementation.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 129
For c = .9999, the restarted iteration has much more trouble. The un-
restarted iteration converges in 10 iterations at a cost of roughly 249 thousand
floating-point operations. The restarted iteration requires 18 iterations and
407 floating-point operations. Even so, both implementations performed bet-
ter than Newton-GMRES.
0 1 2 3 4 5 6
10
-8
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 7.2. Broyden’s method for the H-equation, c = .9.
0 2 4 6 8 10 12 14 16 18
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Iterations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 7.3. Broyden’s method for the H-equation, c = .9999.
One should take some care in drawing general conclusions. While Broyden’s
method on this example performed better than Newton-GMRES, the storage
in Newton-GMRES is used in the inner iteration, while that in Broyden in the
nonlinear iteration. If many nonlinear iterations are needed, it may be the case
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
130 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
that Broyden’s method may need restarts while the GMRES iteration for the
linear problem for the Newton step may not.
Convection-diffusion equation. We consider the nonlinear convection-
diffusion equation from ¸ 6.4.2. We use the same mesh width h = 1/32 on
a uniform square grid. We set C = 20, u
0
= 0, and τ
r
= τ
a
= h
2
. We
preconditioned with the Poisson solver fish2d.
We allowed for an increase in the residual, which happened on the second
iterate. This is not supported by the local convergence theory for nonlinear
problems; however it can be quite successful in practice. One can do this in
brsol by setting the third entry in the parameter list to 1, as one would if the
problem were truly linear. In ¸ 8.4.3 we will return to this example.
In Fig. 7.4 we compare Broyden’s method without restarts (solid line) to
Broyden’s method with restarts every 8 iterates (dashed line). The unrestarted
Broyden iteration converges in 12 iterations at a cost of roughly 2.2 million
floating-point operations. This is a modest improvement over the Newton-
GMRES cost (reported in ¸ 6.4.2) of 2.6 million floating-point operations.
However, the residual norms do not monotonically decrease in the Broyden
iteration (we fix this in ¸ 8.3.2) and more storage was used.
0 2 4 6 8 10 12 14 16
10
-4
10
-3
10
-2
10
-1
10
0
10
1
Iterations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 7.4. Broyden’s method for nonlinear PDE.
At most five inner GMRES iterations were required. Therefore the GMRES
inner iteration for the Newton-GMRES approach needed storage for at most
10 vectors (the Krylov basis, s, x
c
, F(x
c
), F(x
c
+ hv)) to accommodate the
Newton-GMRES iteration. The Broyden iteration when restarted every n
iterates, requires storage of n+2 vectors ¦s
j
¦
n−1
j=0
, x, and z (when stored in the
same place as F(x)). Hence restarting the Broyden iteration every 8 iterations
most closely corresponds to the Newton-GMRES requirements. This approach
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 131
took 15 iterations to terminate at a cost of 2.4 million floating-point operations,
but the convergence was extremely irregular after the restart.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
132 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
7.5. Exercises on Broyden’s method
7.5.1. Prove Proposition 7.2.1.
7.5.2. Prove Proposition 7.3.1.
7.5.3. Extend Proposition 7.3.1 by showing that if A is a nonsingular N N
matrix, U is N M, V is M N, then A + UV is nonsingular if and
only if I + V A
−1
U is a nonsingular M M matrix. If this is the case,
then
(A+UV )
−1
= A
−1
−A
−1
U(I +V A
−1
U)
−1
V A
−1
.
This is the Sherman–Morrison–Woodbury formula [68], [199]. See [102]
for further generalizations.
7.5.4. Duplicate the results reported in ¸ 7.4. Vary the parameters in the
equations and the number of vectors stored by Broyden’s method and
report on their effects on performance. What happens in the differential
equation examples if preconditioning is omitted?
7.5.5. Solve the H-equation with Broyden’s method and c = 1. Set τ
a
= τ
r
=
10
−8
. What q-factor for linear convergence do you see? Can you account
for this by applying the secant method to the equation x
2
= 0? See [53]
for a complete explanation.
7.5.6. Try to solve the nonlinear convection-diffusion equation in ¸ 6.4 with
Broyden’s method using various values for the parameter C. How does
the performance of Broyden’s method differ from Newton-GMRES?
7.5.7. Modify nsol to use Broyden’s method instead of the chord method after
the initial Jacobian evaluation. How does this compare with nsol on the
examples in Chapter 5.
7.5.8. Compare the performance of Broyden’s method and Newton-GMRES on
the Modified Bratu Problem
−∇
2
u +du
x
+e
u
= 0
on (0, 1) (0, 1) with homogeneous Dirichlet boundary conditions.
Precondition with the Poisson solver fish2d. Experiment with various
mesh widths, initial iterates, and values for the convection coefficient d.
7.5.9. The bad Broyden method [26], so named because of its inferior perfor-
mance in practice [63], updates an approximation to the inverse of the
Jacobian at the root so that B
−1
≈ F

(x

)
−1
satisfies the inverse secant
equation
B
−1
+
y = s (7.45)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BROYDEN’S METHOD 133
with the rank-one update
B
−1
+
= B
−1
c
+
(s −B
−1
c
y)y
T
|y|
2
2
.
Show that the bad Broyden method is locally superlinearly convergent
if the standard assumptions hold. Try to implement the bad Broyden
method in a storage-efficient manner using the ideas from [67] and ¸ 7.3.
Does anything go wrong? See [67] for more about this.
7.5.10. The Schubert or sparse Broyden algorithm [172], [27] is a quasi-Newton
update for sparse matrices that enforces both the secant equation and the
sparsity pattern. For problems with tridiagonal Jacobian, for example,
the update is
(B
+
)
ij
= (B
c
)
ij
+
(y −B
c
s)
i
s
j

i+1
k=i−1
s
2
k
for 1 ≤ i, j ≤ N and [i − j[ ≤ 1. Note that only the subdiagonal, su-
perdiagonal, and main diagonal are updated. Read about the algorithm
in [172] and prove local superlinear convergence under the standard as-
sumptions and the additional assumption that the sparsity pattern of B
0
is the same as that of F

(x

). How would the Schubert algorithm be
affected by preconditioning? Compare your analysis with those in [158],
[125], and [189].
7.5.11. Implement the bad Broyden method, apply it to the examples in ¸ 7.4,
and compare the performance to that of Broyden’s method. Discuss the
differences in implementation from Broyden’s method.
7.5.12. Implement the Schubert algorithm on an unpreconditioned discretized
partial differential equation (such as the Bratu problem from Exer-
cise 7.5.8) and compare it to Newton-GMRES and Broyden’s method.
Does the relative performance of the methods change as h is decreased?
This is interesting even in one space dimension where all matrices are
tridiagonal. References [101], [93], [92], and [112], are relevant to this
exercise.
7.5.13. Use your result from Exercise 5.7.14 in Chapter 5 to numerically estimate
the q-order of Broyden’s method for some of the examples in this
section (both linear and nonlinear). What can you conclude from your
observations?
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
134 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Chapter 8
Global Convergence
By a globally convergent algorithm we mean an algorithm with the property
that for any initial iterate the iteration either converges to a root of F or fails
to do so in one of a small number of ways. Of the many such algorithms we
will focus on the class of line search methods and the Armijo rule [3], [88],
implemented inexactly [24], [25], [70].
Other methods, such as trust region [153], [154], [181], [129], [63], [24],
[25], [173], [70], and continuation/homotopy methods [1], [2], [109], [159],
[198], can be used to accomplish the same objective. We select the line search
paradigm because of its simplicity, and because it is trivial to add a line search
to an existing locally convergent implementation of Newton’s method. The
implementation and analysis of the line search method we develop in this
chapter do not depend on whether iterative or direct methods are used to
compute the Newton step or upon how or if the Jacobian is computed and
stored. We begin with single equations, where the algorithms can be motivated
and intuition developed with a very simple example.
8.1. Single equations
If we apply Newton’s method to find the root x

= 0 of the function
f(x) = arctan(x) with initial iterate x
0
= 10 we find that the initial iterate
is too far from the root for the local convergence theory in Chapter 5 to be
applicable. The reason for this is that f

(x) = (1 + x
2
)
−1
is small for large
x; f(x) = arctan(x) ≈ ±π/2, and hence the magnitude of the Newton step
is much larger than that of the iterate itself. We can see this effect in the
sequence of iterates:
10, −138, 2.9 10
4
, −1.5 10
9
, 9.9 10
17
,
a failure of Newton’s method that results from an inaccurate initial iterate.
If we look more closely, we see that the Newton step s = −f(x
c
)/f

(x
c
) is
pointing toward the root in the sense that sx
c
< 0, but the length of s is too
large. This observation motivates the simple fix of reducing the size of the step
until the size of the nonlinear residual is decreased. A prototype algorithm is
given below.
135
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
136 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Algorithm 8.1.1. lines1(x, f, τ)
1. r
0
= [f(x)[
2. Do while [f(x)[ > τ
r
r
0

a
(a) If f

(x) = 0 terminate with failure.
(b) s = −f(x)/f

(x) (search direction)
(c) x
t
= x +s (trial point)
(d) If [f(x
t
)[ < [f(x)[ then x = x
t
(accept the step)
else
s = s/2 goto 2c (reject the step)
When we apply Algorithm lines1 to our simple example, we obtain the
sequence
10, −8.5, 4.9, −3.8, 1.4, −1.3, 1.2, −.99, .56, −0.1, 9 10
−4
, −6 10
−10
.
The quadratic convergence behavior of Newton’s method is only apparent
very late in the iteration. Iterates 1, 2, 3, and 4 required 3, 3, 2, and 2
steplength reductions. After the fourth iterate, decrease in the size of the
nonlinear residual was obtained with a full Newton step. This is typical of the
performance of the algorithms we discuss in this chapter.
We plot the progress of the iteration in Fig. 8.1. We plot
[arctan(x)/arctan(x
0
)[
as a function of the number of function evaluations with the solid line. The
outer iterations are identified with circles as in ¸ 6.4. One can see that most of
the work is in the initial phase of the iteration. In the terminal phase, where
the local theory is valid, the minimum number of two function evaluations per
iterate is required. However, even in this terminal phase rapid convergence
takes place only in the final three outer iterations.
Note the differences between Algorithm lines1 and Algorithm nsol. One
does not set x
+
= x +s without testing the step to see if the absolute value of
the nonlinear residual has been decreased. We call this method a line search
because we search along the line segment
(x
c
, x
c
−f(x
c
)/f

(x
c
))
to find a decrease in [f(x
c
)[. As we move from the right endpoint to the left,
some authors [63] refer to this as backtracking.
Algorithm lines1 almost always works well. In theory, however, there is
the possibility of oscillation about a solution without convergence [63]. To
remedy this and to make analysis possible we replace the test for simple
decrease with one for sufficient decrease. The algorithm is to compute a search
direction d which for us will be the Newton direction
d = −f(x
c
)/f

(x
c
)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 137
0 5 10 15 20 25 30 35
10
-10
10
-9
10
-8
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 8.1. Newton–Armijo iteration for inverse tangent.
and then test steps of the form s = λd, with λ = 2
−j
for some j ≥ 0, until
f(x +s) satisfies
[f(x
c
+λd)[ < (1 −αλ)[f(x
c
)[. (8.1)
The condition in (8.1) is called sufficient decrease of [f[. The parameter
α ∈ (0, 1) is a small, but positive, number intended to make (8.1) as easy
as possible to satisfy. We follow the recent optimization literature [63] and
set α = 10
−4
. Once sufficient decrease has been obtained we accept the step
s = λd. This strategy, from [3] (see also [88]) is called the Armijo rule. Note
that we must now distinguish between the step and the Newton direction,
something we did not have to do in the earlier chapters.
Algorithm 8.1.2. nsola1(x, f, τ)
1. r
0
= [f(x)[
2. Do while [f(x)[ > τ
r
r
0

a
(a) If f

(x) = 0 terminate with failure.
(b) d = −f(x)/f

(x) (search direction)
(c) λ = 1
i. x
t
= x +λd (trial point)
ii. If [f(x
t
)[ < (1 −αλ)[f(x)[ then x = x
t
(accept the step)
else
λ = λ/2 goto 2(c)i (reject the step)
This is essentially the algorithm implemented in the MATLAB code nsola.
We will analyze Algorithm nsola in ¸ 8.2. We remark here that there is
nothing critical about the reduction factor of 2 in the line search. A factor of
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
138 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
10 could well be better in situations in which small values of λ are needed for
several consecutive steps (such as our example with the arctan function). In
that event a reduction of a factor of 8 would require three passes through the
loop in nsola1 but only a single reduction by a factor of 10. This could be
quite important if the search direction were determined by an inexact Newton
method or an approximation to Newton’s method such as the chord method
or Broyden’s method.
On the other hand, reducing λ by too much can be costly as well. Taking
full Newton steps ensures fast local convergence. Taking as large a fraction as
possible helps move the iteration into the terminal phase in which full steps
may be taken and fast convergence expected. We will return to the issue of
reduction in λ in ¸ 8.3.
8.2. Analysis of the Armijo rule
In this section we extend the one-dimensional algorithm in two ways before
proceeding with the analysis. First, we accept any direction that satisfies the
inexact Newton condition (6.1). We write this for the Armijo sequence as
|F

(x
n
)d
n
+F(x
n
)| ≤ η
n
|F(x
n
)|. (8.2)
Our second extension is to allow more flexibility in the reduction of λ. We
allow for any choice that produces a reduction that satisfies
σ
0
λ
old
≤ λ
new
≤ σ
1
λ
old
,
where 0 < σ
0
< σ
1
< 1. One such method, developed in ¸ 8.3, is to minimize
an interpolating polynomial based on the previous trial steps. The danger is
that the minimum may be too near zero to be of much use and, in fact, the
iteration may stagnate as a result. The parameter σ
0
safeguards against that.
Safeguarding is an important aspect of many globally convergent methods and
has been used for many years in the context of the polynomial interpolation
algorithms we discuss in ¸ 8.3.1 [54], [63], [86], [133], [87]. Typical values of σ
0
and σ
1
are .1 and .5. Our algorithm nsola incorporates these ideas.
While (8.2) is motivated by the Newton-iterative paradigm, the analysis
here applies equally to methods that solve the equations for the Newton step
directly (so η
n
= 0), approximate the Newton step by a quasi-Newton method,
or even use the chord method provided that the resulting direction d
n
satisfies
(8.2). In Exercise 8.5.9 you are asked to implement the Armijo rule in the
context of Algorithm nsol.
Algorithm 8.2.1. nsola(x, F, τ, η).
1. r
0
= |F(x)|
2. Do while |F(x)| > τ
r
r
0

a
(a) Find d such that |F

(x)d +F(x)| ≤ η|F(x)|
If no such d can be found, terminate with failure.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 139
(b) λ = 1
i. x
t
= x +λd
ii. If |F(x
t
)| < (1 −αλ)|F(x)| then x = x
t
else
Choose σ ∈ [σ
0
, σ
1
]
λ = σλ
goto 2(b)i
Note that step 2a must allow for the possibility that F

is ill-conditioned
and that no direction can be found that satisfies (8.2). If, for example, step
2a were implemented with a direct factorization method, ill-conditioning of F

might be detected as part of the factorization.
Let ¦x
n
¦ be the iteration produced by Algorithm nsola with initial iterate
x
0
. The algorithm can fail in some obvious ways:
1. F

(x
n
) is singular for some n. The inner iteration could terminate with
a failure.
2. x
n
→ ¯ x, a local minimum of |F| which is not a root.
3. |x
n
| → ∞.
Clearly if F has no roots, the algorithm must fail. We will see that if F has no
roots then either F

(x
n
) has a limit point which is singular or ¦x
n
¦ becomes
unbounded.
The convergence theorem is the remarkable statement that if ¦x
n
¦ and
¦|F

(x
n
)
−1
|¦ are bounded (thereby eliminating premature failure and local
minima of |F| that are not roots) and F

is Lipschitz continuous in a
neighborhood of the entire sequence ¦x
n
¦, then the iteration converges to a
root of F at which the standard assumptions hold, full steps are taken in the
terminal phase, and the convergence is q-quadratic.
We begin with a formal statement of our assumption that F

is uniformly
Lipschitz continuous and bounded away from zero on the sequence ¦x
n
¦.
Assumption 8.2.1. There are r, γ, m
f
> 0 such that F is defined, F

is
Lipschitz continuous with Lipschitz constant γ, and |F

(x)
−1
| ≤ m
f
on the
set
Ω(¦x
n
¦, r) =

_
n=0
¦x[ |x −x
n
| ≤ r¦.
Assumption 8.2.1 implies that the line search phase (step 2b in Algorithm
nsola) will terminate in a number of cycles that is independent of n. We show
this, as do [25] and [70], by giving a uniform lower bound on λ
n
(assuming
that F(x
0
) ,= 0). Note how the safeguard bound σ
0
enters this result. Note
also the very modest conditions on η
n
.
Lemma 8.2.1. Let x
0
∈ R
N
and α ∈ (0, 1) be given. Assume that ¦x
n
¦ is
given by Algorithm nsola with
¦η
n
¦ ⊂ (0, ¯ η] ⊂ (0, 1 −α) (8.3)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
140 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
and that Assumption 8.2.1 holds. Then either F(x
0
) = 0 or
λ
n

¯
λ = σ
0
min
_
r
m
f
|F(x
0
)|
,
2(1 −α − ¯ η)
m
2
f
|F(x
0
)|(1 + ¯ η)
2
γ
_
. (8.4)
Proof. Let n ≥ 0. By the fundamental theorem of calculus and (8.2) we
have, for any λ ∈ [0, 1]
F(x
n
+λd
n
) = F(x
n
) +λF

(x
n
)d
n
+
_
1
0
(F

(x
n
+tλd
n
) −F

(x
n
))λd
n
dt
= (1 −λ)F(x
n
) +λξ
n
+
_
1
0
(F

(x
n
+tλd
n
) −F

(x
n
))λd
n
dt,
where, by the bound η
n
≤ ¯ η,

n
| ≤ η
n
|F(x
n
)| ≤ ¯ η|F(x
n
)|.
The step acceptance condition in nsola implies that ¦|F(x
n
)|¦ is a decreasing
sequence and therefore
|d
n
| = |F

(x
n
)
−1

n
−F(x
n
))| ≤ m
f
(1 + ¯ η)|F(x
n
)|.
Therefore, by Assumption 8.2.1 F

is Lipschitz continuous on the line segment
[x
n
, x
n
+λd
n
] whenever
λ ≤
¯
λ
1
= r/(m
f
|F(x
0
)|).
Lipschitz continuity of F

implies that if λ ≤
¯
λ
1
then
|F(x
n
+λd
n
)| ≤ (1 −λ)|F(x
n
)| +λ¯ η|F(x
n
)| +
γ
2
λ
2
|d
n
|
2
≤ (1 −λ)|F(x
n
)[ +λ¯ η|F(x
n
)| +
m
2
f
γ(1 + ¯ η)
2
λ
2
|F(x
n
)|
2
2
≤ (1 −λ + ¯ ηλ)|F(x
n
)| +λ|F(x
n
)|
m
2
f
γλ(1 + ¯ η)
2
|F(x
0
)|
2
< (1 −αλ)|F(x
n
)|,
which will be implied by
λ ≤
¯
λ
2
= min
_
¯
λ
1
,
2(1 −α − ¯ η)
(1 + ¯ η)
2
m
2
f
γ|F(x
0
)|
_
.
This shows that λ can be no smaller than σ
0
¯
λ
2
, which completes the proof.
Now we can show directly that the sequence of Armijo iterates either is
unbounded or converges to a solution. Theorem 8.2.1 says even more. The
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 141
sequence converges to a root at which the standard assumptions hold, so in the
terminal phase of the iteration the step lengths are all 1 and the convergence
is q-quadratic.
Theorem 8.2.1. Let x
0
∈ R
N
and α ∈ (0, 1) be given. Assume that ¦x
n
¦
is given by Algorithm nsola, ¦η
n
¦ satisfies (8.3), ¦x
n
¦ is bounded, and that
Assumption 8.2.1 holds. Then ¦x
n
¦ converges to a root x

of F at which the
standard assumptions hold, full steps are taken for n sufficiently large, and the
convergence behavior in the final phase of the iteration is that given by the local
theory for inexact Newton methods (Theorem 6.1.2).
Proof. If F(x
n
) = 0 for some n, then we are done because Assump-
tion 8.2.1 implies that the standard assumptions hold at x

= x
n
. Otherwise
Lemma 8.2.1 implies that F(x
n
) converges q-linearly to zero with q-factor at
most (1 −α
¯
λ).
The boundedness of the sequence ¦x
n
¦ implies that a subsequence ¦x
n
k
¦
converges, say to x

. Since F is continuous, F(x

) = 0. Eventually
[x
n
k
− x

[ < r, where r is the radius in Assumption 8.2.1 and therefore the
standard assumptions hold at x

.
Since the standard assumptions hold at x

, there is δ such that if x ∈ B(δ),
the Newton iteration with x
0
= x will remain in B(δ) and converge q-
quadratically to x

. Hence as soon as x
n
k
∈ B(δ), the entire sequence, not
just the subsequence, will remain in B(δ) and converge to x

. Moreover,
Theorem 6.1.2 applies and hence full steps can be taken.
At this stage we have shown that if the Armijo iteration fails to converge
to a root either the continuity properties of F or nonsingularity of F

break
down as the iteration progresses (Assumption 8.2.1 is violated) or the iteration
becomes unbounded. This is all that one could ask for in terms of robustness
of such an algorithm.
Thus far we have used for d the inexact Newton direction. It could well be
advantageous to use a chord or Broyden direction, for example. All that one
needs to make the theorems and proofs in ¸ 8.2 hold is (8.2), which is easy to
verify, in principle, via a forward difference approximation to F

(x
n
)d.
8.3. Implementation of the Armijo rule
The MATLAB code nsola is a modification of nsolgm which provides for a
choice of several Krylov methods for computing an inexact Newton direction
and globalizes Newton’s method with the Armijo rule. We use the l
2
norm in
that code and in the numerical results reported in ¸ 8.4. If GMRES is used
as the linear solver, then the storage requirements are roughly the same as for
Algorithm nsolgm if one reuses the storage location for the finite difference
directional derivative to test for sufficient decrease.
In Exercise 8.5.9 you are asked to do this with nsol and, with each
iteration, numerically determine if the chord direction satisfies the inexact
Newton condition. In this case the storage requirements are roughly the same
as for Algorithm nsol. The iteration can be very costly if the Jacobian must be
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
142 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
evaluated and factored many times during the phase of the iteration in which
the approximate solution is far from the root.
We consider two topics in this section. The first, polynomial interpolation
as a way to choose λ, is applicable to all methods. The second, how Broyden’s
method must be implemented, is again based on [67] and is more specialized.
Our Broyden–Armijo code does not explicitly check the inexact Newton
condition and thereby saves an evaluation of F. Evaluation of F

(x
c
)d
c
if the
full step is not accepted (here d
c
is the Broyden direction) would not only verify
the inexact Newton condition but provide enough information for a two-point
parabolic line search. The reader is asked to pursue this in Exercise 8.5.10.
8.3.1. Polynomial line searches. In this section we consider an elabora-
tion of the simple line search scheme of reduction of the steplength by a fixed
factor. The motivation for this is that some problems respond well to one or
two reductions in the steplength by modest amounts (such as 1/2) and others
require many such reductions, but might respond well to a more aggressive
steplength reduction (by factors of 1/10, say). If one can model the decrease
in |F|
2
as the steplength is reduced, one might expect to be able to better
estimate the appropriate reduction factor. In practice such methods usually
perform better than constant reduction factors.
If we have rejected k steps we have in hand the values
|F(x
n
)|
2
, |F(x
n

1
d
n
)|
2
, . . . |F(x
n

k−1
d
n
)|
2
.
We can use this iteration history to model the scalar function
f(λ) = |F(x
n
+λd
n
)|
2
2
with a polynomial and use the minimum of that polynomial as the next
steplength. We consider two ways of doing this that use second degree
polynomials which we compute using previously computed information.
After λ
c
has been rejected and a model polynomial computed, we compute
the minimum λ
t
of that polynomial analytically and set
λ
+
=
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
σ
0
λ
c
if λ
t
< σ
0
λ
c
,
σ
1
λ
c
if λ
t
> σ
1
λ
c
,
λ
t
otherwise.
(8.5)
Two-point parabolic model. In this approach we use values of f and f

at λ = 0 and the value of f at the current value of λ to construct a 2nd degree
interpolating polynomial for f.
f(0) = |F(x
n
)|
2
2
is known. Clearly f

(0) can be computed as
f

(0) = 2(F

(x
n
)
T
d
n
)
T
F(x
n
) = 2F(x
n
)
T
(F

(x
n
)d
n
). (8.6)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 143
Use of (8.6) requires evaluation of F

(x
n
)d
n
, which can be obtained by
examination of the final residual from GMRES or by expending an additional
function evaluation to compute F

(x
n
)d
n
with a forward difference. In any
case, our approximation to f

(0) should be negative. If it is not, it may
be necessary to compute a new search direction. This is a possibility with
directions other than inexact Newton directions, such as the Broyden direction.
The polynomial p(λ) such that p, p

agree with f, f

at 0 and p(λ
c
) = f(λ
c
)
is
p(λ) = f(0) +f

(0)λ +cλ
2
,
where
c =
f(λ
c
) −f(0) −f

(0)λ
c
λ
2
c
.
Our approximation of f

(0) < 0, so if f(λ
c
) > f(0), then c > 0 and p(λ) has a
minimum at
λ
t
= −f

(0)/(2c) > 0.
We then compute λ
+
with (8.5).
Three-point parabolic model. An alternative to the two-point model that
avoids the need to approximate f

(0) is a three-point model, which uses f(0)
and the two most recently rejected steps to create the parabola. The MATLAB
code parab3p implements the three-point parabolic line search and is called
by the two codes nsola and brsola.
In this approach one evaluates f(0) and f(1) as before. If the full step is
rejected, we set λ = σ
1
and try again. After the second step is rejected, we
have the values
f(0), f(λ
c
), and f(λ

),
where λ
c
and λ

are the most recently rejected values of λ. The polynomial
that interpolates f at 0, λ
c
, λ

is
p(λ) = f(0) +
λ
λ
c
−λ

_
(λ −λ

)(f(λ
c
) −f(0))
λ
c
+

c
−λ)(f(λ

) −f(0))
λ

_
.
We must consider two situations. If
p

(0) =
2
λ
c
λ


c
−λ

)


(f(λ
c
) −f(0)) −λ
c
(f(λ

) −f(0)))
is positive, then we set λ
t
to the minimum of p
λ
t
= −p

(0)/p

(0)
and apply the safeguarding step (8.5) to compute λ
+
. If p

(0) ≤ 0 one could
either set λ
+
to be the minimum of p on the interval [σ
0
λ, σ
1
λ] or reject the
parabolic model and simply set λ
+
to σ
0
λ
c
or σ
1
λ
c
. We take the latter approach
and set λ
+
= σ
1
λ
c
. In the MATLAB code nsola from the collection, we
implement this method with σ
0
= .1, σ
1
= .5.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
144 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
Interpolation with a cubic polynomial has also been suggested [87], [86],
[63], [133]. Clearly, the polynomial modeling approach is heuristic. The
safeguarding and Theorem 8.2.1 provide insurance that progress will be made
in reducing |F|.
8.3.2. Broyden’s method. In Chapter 7 we implemented Broyden’s
method at a storage cost of one vector for each iteration. The relation
y − B
c
s = F(x
+
) was not critical to this and we may also incorporate a line
search at a cost of a bit of complexity in the implementation. As in ¸ 7.3, our
approach is a nonlinear version of the method in [67].
The difference from the development in ¸ 7.3 is that the simple relation
between the sequence ¦w
n
¦ and ¦v
n
¦ is changed. If a line search is used, then
s
n
= −λ
n
B
−1
n
F(x
n
)
and hence
y
n
−B
n
s
n
= F(x
n+1
) −(1 −λ
n
)F(x
n
). (8.7)
If, using the notation in ¸ 7.3, we set
u
n
=
y
n
−B
n
s
n
|s
n
|
2
, v
n
=
s
n
|s
n
|
2
, and w
n
= (B
−1
n
u
n
)/(1 +v
T
n
B
−1
n
u
n
),
we can use (8.7) and (7.38) to obtain
w
n
= λ
n
_
−d
n+1
|s
n
|
2
+ (λ
−1
n
−1)
s
n
|s
n
|
2
_
= λ
n
_
−s
n+1

n+1
|s
n
|
2
+ (λ
−1
n
−1)
s
n
|s
n
|
2
_
,
(8.8)
where
d
n+1
= −B
−1
n+1
F(x
n+1
)
is the search direction used to compute x
n+1
. Note that (8.8) reduces to (7.41)
when λ
n
= 1 and λ
n+1
= 1 (so d
n+1
= s
n+1
).
We can use the first equality in (8.8) and the relation
d
n+1
= −
_
I −
w
n
s
T
n
|s
n
|
2
_
B
−1
n
F(x
n+1
)
to solve for d
n+1
and obtain an analog of (7.44)
d
n+1
= −
|s
n
|
2
2
B
−1
n
F(x
n+1
) −(1 −λ
n
)s
T
n
B
−1
n
F(x
n+1
)s
n
|s
n
|
2
2

n
s
T
n
B
−1
n
F(x
n+1
)
(8.9)
We then use the second equality in (8.8) to construct B
−1
c
F(x
+
) as we did in
Algorithm brsol. Verification of (8.8) and (8.9) is left to Exercise 8.5.5.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 145
Algorthm brsola uses these ideas. We do not include details of the line
search or provide for restarts or a limited memory implementation. The
MATLAB implementation, provided in the collection of MATLAB codes, uses
a three-point parabolic line search and performs a restart when storage is
exhausted like Algorithm brsol does.
It is possible that the Broyden direction fails to satisfy (8.2). In fact, there
is no guarantee that the Broyden direction is a direction of decrease for |F|
2
.
The MATLAB code returns with an error message if more than a fixed number
of reductions in the step length are needed. Another approach would be to
compute F

(x
c
)d numerically if a full step is rejected and then test (8.2) before
beginning the line search. As a side effect, an approximation of f

(0) can be
obtained at no additional cost and a parabolic line search based on f(0), f

(0),
and f(1) can be used. In Exercise 8.5.10 the reader is asked to fully develop
and implement these ideas.
Algorithm 8.3.1. brsola(x, F, τ, maxit, nmax)
1. Initialize:
F
c
= F(x) r
0
= |F(x)|
2
, n = −1,
d = −F(x), compute λ
0
, s
0
= λ
0
d
2. Do while n ≤ maxit
(a) n = n + 1
(b) x = x +s
n
(c) Evaluate F(x)
(d) If |F(x)|
2
≤ τ
r
r
0

a
exit.
(e) i. z = −F(x)
ii. for j = 0, n −1
a = −λ
j

j+1
, b = 1 −λ
j
z = z + (as
j+1
+bs
j
)s
T
j
z/|s
j
|
2
2
(f) d = (|s
n
|
2
2
z + (1 −λ
n
)s
T
n
zs
n
)/(|s
n
|
2
2
−λ
n
s
T
n
z)
(g) Compute λ
n+1
with a line search.
(h) Set s
n+1
= λ
n+1
d, store |s
n+1
|
2
and λ
n+1
.
Since d
n+1
and s
n+1
can occupy the same location, the storage require-
ments of Algorithm brsola would appear to be essentially the same as those
of brsol. However, computation of the step length requires storage of both the
current point x and the trial point x + λs before a step can be accepted. So,
the storage of the globalized methods exceeds that of the local method by one
vector. The MATLAB implementation of brsola in the collection illustrates
this point.
8.4. Examples for Newton–Armijo
The MATLAB code nsola is a modification of nsolgm that incorporates the
three-point parabolic line search from ¸ 8.3.1 and also changes η using (6.20)
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
146 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
once the iteration is near the root. We compare this strategy, using GMRES
as the linear solver, with the constant η method as we did in ¸ 6.4 with γ = .9.
As before, in all the figures we plot |F(x
n
)|
2
/|F(x
0
)|
2
against the number
of function evaluations required by all line searches and inner and outer
iterations to compute x
n
. Counts of function evaluations corresponding to
outer iterations are indicated by circles. We base our absolute error criterion
on the norm | |
2
/

N as we did in ¸ 6.4.2 and 7.4.
In ¸ 8.4.3 we compare the Broyden–Armijo method with Newton-GMRES
for those problems for which both methods were successful. We caution the
reader now, and will repeat the caution later, that if the initial iterate is far
from the solution, an inexact Newton method such as Newton-GMRES can
succeed in many case in which a quasi-Newton method can fail because the
quasi-Newton direction may not be an inexact Newton direction. However,
when both methods succeed, the quasi-Newton method, which requires a single
function evaluation for each outer iteration when full steps can be taken, may
well be most efficient.
8.4.1. Inverse tangent function. Since we have easy access to analytic
derivatives in this example, we can use the two-point parabolic line search.
We compare the two-point parabolic line search with the constant reduction
search (σ
0
= σ
1
= .5) for the arctan function. In Fig. 8.2 we plot the iteration
history corresponding to the parabolic line search with the solid line and that
for the constant reduction with the dashed line. We use x
0
= 10 as the initial
iterate with τ
r
= τ
a
= 10
−8
. The parabolic line search required 7 outer iterates
and 21 function evaluations in contrast with the constant reduction searches 11
outer iterates and 33 function evaluations. In the first outer iteration, both line
searches take three steplength reductions. However, the parabolic line search
takes only one reduction in the next three iterations and none thereafter. The
constant reduction line search took three reductions in the first two outer
iterations and two each in following two.
8.4.2. Convection-diffusion equation. We consider a much more diffi-
cult problem. We take (6.21) from ¸ 6.4.2,
−∇
2
u +Cu(u
x
+u
y
) = f
with the same right hand side f, initial iterate u
0
= 0, 31 31 mesh and
homogeneous Dirichlet boundary conditions on the unit square (0, 1) (0, 1)
that we considered before. Here, however, we set C = 100. This makes a
globalization strategy critical (Try it without one!). We set τ
r
= τ
a
= h
2
/10.
This is a tighter tolerance that in ¸ 6.4.2 and, because of the large value of C
and resulting ill-conditioning, is needed to obtain an accurate solution.
In Fig. 8.3 we show the progress of the iteration for the unpreconditioned
equation. For this problem we plot the progress of the iteration using η = .25
with the solid line and using the sequence given by (6.20) with the dashed
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 147
0 5 10 15 20 25 30 35
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 8.2. Newton–Armijo for the arctan function.
line. We set the parameters in (6.20) to γ = .9 and η
max
= .25. This problem
is very poorly conditioned and much tighter control on the inner iteration is
needed than for the other problems. The two approaches to selection of the
forcing term performed very similarly. The iterations required 75.3 (constant
η) and 75.4 (varying η) million floating point operations, 759 (constant η) and
744 (varying η) function evaluations, and 25 (constant η) and 22 (varying η)
outer iterations.
0 100 200 300 400 500 600 700 800
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 8.3. Newton–Armijo for the PDE, C = 100.
We consider the preconditioned problem in Fig. 8.4. In the computation we
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
148 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
preconditioned (6.21) with the fast Poisson solver fish2d. In Fig. 8.4 we show
the progress of the iteration for the preconditioned equation. For this problem
we plot the progress of the iteration using η = .25 with the solid line and using
the sequence given by (6.20) with the dashed line. We set the parameters in
(6.20) to γ = .9 and η
max
= .99.
The constant η iteration terminated after 79 function evaluations, 9 outer
iterations, and roughly 13.5 million floating-point operations. The line search
in this case reduced the step length three times on the first iteration and twice
on the second and third.
The iteration in which ¦η
n
¦ is given by (6.20) terminated after 70 function
evaluations, 9 outer iterations, and 12.3 million floating-point operations. The
line search in the non-constant η case reduced the step length three times on
the first iteration and once on the second. The most efficient iteration, with
the forcing term given by (6.20), required at most 16 inner iterations while the
constant η approach needed at most 10.
0 10 20 30 40 50 60 70 80
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 8.4. Newton–Armijo for the PDE, C = 100.
8.4.3. Broyden–Armijo. We found that the Broyden–Armijo line search
failed on all the unpreconditioned convection diffusion equation examples. In
these cases the Jacobian is not well approximated by a low rank perturbation
of the identity [112] and the ability of the inexact Newton iteration to find a
good search direction was the important factor.
We begin by revisiting the example in ¸ 6.4.2 with C = 20. As reported
in ¸ 6.4.2, the Newton-GMRES iteration always took full steps, terminating
in 4 outer iterations, 16 function evaluations, and 2.6 million floating-point
operations. When compared to the Broyden’s method results in ¸ 6.4.2,
when increases in the residual were allowed, the Broyden–Armijo costs reflect
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 149
an improvement in efficiency. The Broyden iteration in ¸ 6.4.2 took 12
iterations and 2.8 million floating-point operations while the Broyden–Armijo
took 9 iterations and required 2 million floating-point operations. We set
τ
a
= τ
r
= h
2
, η
max
= .5, and γ = .9 as we did in ¸ 6.4.2.
In our implementation of brsola the three-point parabolic line search was
used. The steplength was reduced on the second iteration (twice) and the third
(once). Figure 8.5 compares the Broyden–Armijo iteration (solid line) to the
GMRES iteration (dashed line).
0 2 4 6 8 10 12 14 16
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 8.5. Newton-GMRES and Broyden–Armijo for the PDE, C = 20.
For the more difficult problem with C = 100, the performance of the
two methods is more similar. In Fig. 8.6 we compare our implementation
of Newton–GMRES–Armijo (dashed line) to Broyden–Armijo (solid line). We
set τ
a
= τ
r
= h
2
/10, η
max
= .99, and γ = .9. The Broyden–Armijo iteration
required 34 nonlinear iterations, roughly 16 million floating-point operations,
and 85 function evaluations. In terms of storage, the Broyden–Armijo iteration
required storage for 37 vectors while the Newton-GMRES iteration needed at
most 16 inner iterations, needing to store 21 vectors, and therefore was much
more efficient in terms of storage.
Restarting the Broyden iteration every 19 iterations would equalize the
storage requirements with Newton–GMRES–Armijo. Broyden’s method suf-
fers under these conditions as one can see from the comparison of Broyden–
Armijo (solid line) and Newton–GMRES–Armijo (dashed line) in Fig. 8.7. The
restarted Broyden–Armijo iteration took 19.5 million floating-point operations,
42 outer iterates, and 123 function evaluations.
From these examples we see that Broyden–Armijo can be very efficient pro-
vided only a small number of iterations are needed. The storage requirements
increase as the nonlinear iteration progresses and this fact puts Broyden’s
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
150 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
0 10 20 30 40 50 60 70 80 90
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 8.6. Newton-GMRES and Broyden–Armijo for the PDE, C = 100.
0 20 40 60 80 100 120 140
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Function Evaluations
R
e
l
a
t
i
v
e

N
o
n
l
i
n
e
a
r

R
e
s
i
d
u
a
l
Fig. 8.7. Newton-GMRES and restarted Broyden–Armijo for the PDE, C = 100.
method at a disadvantage to Newton-GMRES when the number of nonlinear
iterations is large.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
GLOBAL CONVERGENCE 151
8.5. Exercises on global convergence
8.5.1. How does the method proposed in [10] differ from the one implemented in
nsola? What could be advantages and disadvantages of that approach?
8.5.2. In what ways are the results in [25] and [70] more general than those in
this section? What ideas do these papers share with the analysis in this
section and with [3] and [88]?
8.5.3. Implement the Newton–Armijo method for single equations. Apply your
code to the following functions with x
0
= 10. Explain your results.
1. f(x) = arctan(x) (i.e., Duplicate the results in ¸ 8.1.)
2. f(x) = arctan(x
2
)
3. f(x) = .9 + arctan(x)
4. f(x) = x(1 + sin(1/x))
5. f(x) = e
x
6. f(x) = 2 + sin(x)
7. f(x) = 1 +x
2
8.5.4. A numerical analyst buys a German sports car for $50,000. He puts
$10,000 down and takes a 7 year installment loan to pay the balance. If
the monthly payments are $713.40, what is the interest rate? Assume
monthly compounding.
8.5.5. Verify (8.8) and (8.9).
8.5.6. Show that if F

is Lipschitz continuous and the iteration ¦x
n
¦ produced
by Algorithm nsola converges to x

with F(x

) ,= 0, then F

(x

) is
singular.
8.5.7. Use nsola to duplicate the results in ¸ 8.4.2. Vary the convection
coefficient in the convection-diffusion equation and the mesh size and
report the results.
8.5.8. Experiment with other linear solvers such as GMRES(m), Bi-CGSTAB,
and TFQMR. This is interesting in the locally convergent case as well.
You might use the MATLAB code nsola to do this.
8.5.9. Modify nsol, the hybrid Newton algorithm from Chapter 5, to use the
Armijo rule. Try to do it in such a way that the chord direction is used
whenever possible.
8.5.10. Modify brsola to test the Broyden direction for the descent property
and use a two-point parabolic line search. What could you do if the
Broyden direction is not a descent direction? Apply your code to the
examples in ¸ 8.4.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
152 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS
8.5.11. Modify nsola to use a cubic polynomial and constant reduction line
searches instead of the quadratic polynomial line search. Compare the
results with the examples in ¸ 8.4.
8.5.12. Does the secant method for equations in one variable always give a
direction that satisfies (8.2) with η
n
bounded away from 1? If not, when
does it? How would you implement a secant-Armijo method in such a
way that the convergence theorem 8.2.1 is applicable?
8.5.13. Under what conditions will the iteration given by nsola converge to a
root x

that is independent of the initial iterate?
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Bibliography
[1] E. L. Allgower and K. Georg, Simplicial and continuation methods for
approximating fixed points and solutions to systems of equations, SIAM Rev., 22
(1980), pp. 28–85.
[2] , Numerical Continuation Methods : An Introduction, Springer-Verlag, New
York, 1990.
[3] L. Armijo, Minimization of functions having Lipschitz-continuous first partial
derivatives, Pacific J. Math., 16 (1966), pp. 1–3.
[4] W. E. Arnoldi, The principle of minimized iterations in the solution of the
matrix eigenvalue problem, Quart. Appl. Math., 9 (1951), pp. 17–29.
[5] S. F. Ashby, T. A. Manteuffel, and J. S. Otto, A comparison of
adaptive Chebyshev and least squares polynomial preconditioning for Hermetian
positive definite linear systems, SIAM J. Sci. Statist. Comput., 13 (1992), pp. 1–
29.
[6] K. E. Atkinson, Iterative variants of the Nystr¨ om method for the numerical
solution of integral equations, Numer. Math., 22 (1973), pp. 17–31.
[7] , An Introduction to Numerical Analysis, 2nd. ed., John Wiley, New York,
1989.
[8] O. Axelsson, Iterative Solution Methods, Cambridge University Press, Cam-
bridge, 1994.
[9] S. Banach, Sur les op´erations dans les ensembles abstraits et leur applications
aux ´equations int´egrales, Fund. Math, 3 (1922), pp. 133–181.
[10] R. E. Bank and D. J. Rose, Global approximate Newton methods, Numer.
Math., 37 (1981), pp. 279–295.
[11] M. S. Barlett, An inverse matrix adjustment arising in discriminant analysis,
Ann. Math. Statist., 22 (1951), pp. 107–111.
[12] R. Barrett, M. Berry, T. F. Chan, J. Demmel, J. Donato, J. Don-
garra, V. Eijkhout, R. Pozo, C. Romine, and H. van der Vorst, Templates
for the Solution of Linear Systems: Building Blocks for Iterative Methods, Society
for Industrial and Applied Mathematics, Philadelphia, PA, 1993.
[13] N. Bi´ cani´ c and K. H. Johnson, Who was ‘-Raphson’?, Internat. J. Numer.
Methods. Engrg., 14 (1979), pp. 148–152.
[14] P. B. Bosma and W. A. DeRooij, Efficient methods to calculate Chan-
drasekhar’s H-functions, Astron. Astrophys., 126 (1983), p. 283.
[15] H. Brakhage,
¨
Uber die numerische Behandlung von Integralgleichungen nach
der Quadraturformelmethode, Numer. Math., 2 (1960), pp. 183–196.
[16] K. E. Brenan, S. L. Campbell, and L. R. Petzold, Numerical Solution
153
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
154 BIBLIOGRAPHY
of Initial Value Problems in Differential-Algebraic Equations, no. 14 in Classics in
Applied Mathematics, SIAM, Philadelphia, 1996.
[17] R. Brent, Algorithms for Minimization Without Deriviatives, Prentice-Hall,
Englewood Cliffs, NJ, 1973.
[18] , Some efficient algorithms for solving systems of nonlinear equations, SIAM
J. Numer. Anal., 10 (1973), pp. 327–344.
[19] W. Briggs, A Multigrid Tutorial, Society for Industrial and Applied Mathemat-
ics, Philadelphia, PA, 1987.
[20] P. N. Brown, A local convergence theory for combined inexact–Newton/ finite–
difference projection methods, SIAM J. Numer. Anal., 24 (1987), pp. 407–434.
[21] P. N. Brown, G. D. Byrne, and A. C. Hindmarsh, VODE: A variable
coefficient ode solver, SIAM J. Sci. Statist. Comput., 10 (1989), pp. 1038–1051.
[22] P. N. Brown and A. C. Hindmarsh, Reduced storage matrix methods in stiff
ODE systems, J. Appl. Math. Comput., 31 (1989), pp. 40–91.
[23] P. N. Brown, A. C. Hindmarsh, and L. R. Petzold, Using Krylov methods
in the solution of large-scale differential-algebraic systems, SIAM J. Sci. Comput., 15
(1994), pp. 1467–1488.
[24] P. N. Brown and Y. Saad, Hybrid Krylov methods for nonlinear systems of
equations, SIAM J. Sci. Statist. Comput., 11 (1990), pp. 450–481.
[25] , Convergence theory of nonlinear Newton-Krylov algorithms, SIAM J.
Optim., 4 (1994), pp. 297–330.
[26] C. G. Broyden, A class of methods for solving nonlinear simultaneous equa-
tions, Math. Comput., 19 (1965), pp. 577–593.
[27] , The convergence of an algorithm for solving sparse nonlinear systems,
Math. Comput., 25 (1971), pp. 285–294.
[28] C. G. Broyden, J. E. Dennis, and J. J. Mor´ e, On the local and superlinear
convergence of quasi-Newton methods, J. Inst. Maths. Applic., 12 (1973), pp. 223–
246.
[29] W. Burmeister, Zur Konvergenz einiger verfahren der konjugierten Richtun-
gen, in Proceedings of Internationaler Kongreß ¨ uber Anwendung der Mathematik in
dem Ingenieurwissenschaften, Weimar, 1975.
[30] I. W. Busbridge, The Mathematics of Radiative Transfer, Cambridge Tracts,
No. 50, Cambridge Univ. Press, Cambridge, 1960.
[31] R. H. Byrd, J. Nocedal, and R. B. Schnabel, Representation of quasi-
Newton matrices and their use in limited memory methods, Math. Programming, 63
(1994), pp. 129–156.
[32] X.-C. Cai, W. D. Gropp, D. E. Keyes, and M. D. Tidriri, Newton-Krylov-
Schwarz methods in CFD, in Proceedings of the International Workshop on the
Navier-Stokes Equations, R. Rannacher, ed., Notes in Numerical Fluid Mechanics,
Braunschweig, 1994, Vieweg Verlag.
[33] S. L. Campbell, I. C. F. Ipsen, C. T. Kelley, and C. D. Meyer, GMRES
and the minimal polynomial, Tech. Report CRSC-TR94-10, North Carolina State
University, Center for Research in Scientific Computation, July 1994. BIT, to appear.
[34] S. L. Campbell, I. C. F. Ipsen, C. T. Kelley, C. D. Meyer, and Z. Q.
Xue, Convergence estimates for solution of integral equations with GMRES, Tech.
Report CRSC-TR95-13, North Carolina State University, Center for Research in
Scientific Computation, March 1995. Journal of Integral Equations and Applications,
to appear.
[35] R. Cavanaugh, Difference Equations and Iterative Processes, PhD thesis,
University of Maryland, 1970.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BIBLIOGRAPHY 155
[36] F. Chaitin-Chatelin, Is nonnormality a serious difficulty ?, Tech. Report
TR/PA/94/18, CERFACS, December 1994.
[37] T. Chan, E. Gallopoulos, V. Simoncini, T. Szeto, and C. Tong, A quasi-
minimal residual variant of the Bi-CGSTAB algorithm for nonsymmetric systems,
SIAM J. Sci. Comput., 15 (1994), p. 338.
[38] T. Chan, R. Glowinski, J. P´ eriaux, and O. Widlund, eds., Domain
Decomposition Methods,Proceedings of the Second International Symposium on
Domain Decomposition Methods, Los Angeles, CA, January 14–16, 1988; Society
for Industrial and Applied Mathematics, Philadelphia, PA, 1989.
[39] , eds., Domain Decomposition Methods, Proceedings of the Third Interna-
tional Symposium on Domain Decomposition Methods, Houston, TX, 1989; Society
for Industrial and Applied Mathematics, Philadelphia, PA, 1990.
[40] , eds., Domain Decomposition Methods, Proceedings of the Fourth Interna-
tional Symposium on Domain Decomposition Methods, Moscow, USSR, 1990; Soci-
ety for Industrial and Applied Mathematics, Philadelphia, PA, 1991.,
[41] S. Chandrasekhar, Radiative Transfer, Dover, New York, 1960.
[42] T. F. Coleman and J. J. Mor´ e, Estimation of sparse Jacobian matrices and
graph coloring problems, SIAM J. Numer. Anal., 20 (1983), pp. 187–209.
[43] T. F. Coleman and C. VanLoan, Handbook for Matrix Computations,
Frontiers in Appl. Math., No. 4, Society for Industrial and Applied Mathematics,
Philadelphia, PA, 1988.
[44] P. Concus, G. H. Golub, and G. Meurant, Block preconditioning for the
conjugate gradient method, SIAM J. Sci. Statist. Comput., 6 (1985), pp. 220–252.
[45] P. Concus, G. H. Golub, and D. P. O’Leary, A generalized conjugate
gradient method for the numerical solution of elliptic partial differential equations,
in Sparse Matrix Computations, J. R. Bunch and D. J. Rose, eds., Academic Press,
1976, pp. 309–332.
[46] E. J. Craig, The N-step iteration procedures, J. Math. Phys., 34 (1955), pp. 64–
73.
[47] A. R. Curtis, M. J. D. Powell, and J. K. Reid, On the estimation of sparse
Jacobian matrices, J. Inst. Math. Appl., 13 (1974), pp. 117–119.
[48] J. W. Daniel, The conjugate gradient method for linear and nonlinear operator
equations, SIAM J. Numer. Anal., 4 (1967), pp. 10–26.
[49] D. W. Decker, H. B. Keller, and C. T. Kelley, Convergence rates for
Newton’s method at singular points, SIAM J. Numer. Anal., 20 (1983), pp. 296–314.
[50] D. W. Decker and C. T. Kelley, Newton’s method at singular points I, SIAM
J. Numer. Anal., 17 (1980), pp. 66–70.
[51] , Convergence acceleration for Newton’s method at singular points, SIAM J.
Numer. Anal., 19 (1982), pp. 219–229.
[52] , Sublinear convergence of the chord method at singular points, Numer.
Math., 42 (1983), pp. 147–154.
[53] , Broyden’s method for a class of problems having singular Jacobian at the
root, SIAM J. Numer. Anal., 22 (1985), pp. 566–574.
[54] T. J. Dekker, Finding a zero by means of successive linear interpolation, in
Constructive Aspects of the Fundamental Theorem of Algebra, P. Henrici, ed., 1969,
pp. 37–48.
[55] R. Dembo, S. Eisenstat, and T. Steihaug, Inexact Newton methods, SIAM
J. Numer. Anal., 19 (1982), pp. 400–408.
[56] R. Dembo and T. Steihaug, Truncated Newton algorithms for large-scale
optimization, Math. Programming, 26 (1983), pp. 190–212.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
156 BIBLIOGRAPHY
[57] J. E. Dennis, On the Kantorovich hypothesis for Newton’s method, SIAM J.
Numer. Anal., 6 (1969), pp. 493–507.
[58] , Toward a unified convergence theory for Newton-like methods, in Nonlinear
Functional Analysis and Applications, L. B. Rall, ed., Academic Press, New York,
1971, pp. 425–472.
[59] J. E. Dennis, J. M. Martinez, and X. Zhang, Triangular decomposition
methods for solving reducible nonlinear systems of equations, SIAM J. Optim., 4
(1994), pp. 358–382.
[60] J. E. Dennis and J. J. Mor´ e, A characterization of superlinear convergence
and its application to quasi-Newton methods, Math. Comput., 28 (1974), pp. 549–560.
[61] , Quasi-Newton methods, methods, motivation and theory, SIAM Rev., 19
(1977), pp. 46–89.
[62] J. E. Dennis and R. B. Schnabel, Least change secant updates for quasi-
Newton methods, SIAM Rev., 21 (1979), pp. 443–459.
[63] , Numerical Methods for Unconstrained Optimization and Nonlinear Equa-
tions, no. 16 in Classics in Applied Mathematics, SIAM, Philadelphia, 1996.
[64] J. E. Dennis and H. F. Walker, Convergence theorems for least change secant
update methods, SIAM J. Numer. Anal., 18 (1981), pp. 949–987.
[65] , Inaccuracy in quasi-Newton methods: Local improvement theorems, in
Mathematical Programming Study 22: Mathematical programming at Oberwolfach
II, North–Holland, Amsterdam, 1984, pp. 70–85.
[66] , Least-change sparse secant updates with inaccurate secant conditions,
SIAM J. Numer. Anal., 22 (1985), pp. 760–778.
[67] P. Deuflhard, R. W. Freund, and A. Walter, Fast secant methods for
the iterative solution of large nonsymmetric linear systems, Impact of Computing in
Science and Engineering, 2 (1990), pp. 244–276.
[68] W. J. Duncan, Some devices for the solution of large sets of simultaneous
linear equations (with an appendix on the reciprocation of partitioned matrices), The
London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science,
Seventh Series, 35 (1944), pp. 660–670.
[69] S. C. Eisenstat and H. F. Walker, Choosing the forcing terms in an inexact
Newton method, SIAM J. Sci. Comput., 17 (1996), pp. 16–32.
[70] , Globally convergent inexact Newton methods, SIAM J. Optim., 4 (1994),
pp. 393–422.
[71] H. C. Elman, Iterative Methods for Large, Sparse, Nonsymmetric Systems of
Linear Equations, PhD thesis, Yale University, New Haven, CT, 1982.
[72] H. C. Elman, Y. Saad, and P. E. Saylor, A hybrid Chebyshev-Krylov
subspace algorithm for solving nonsymmetric systems of linear equations, SIAM J.
Sci. Statist. Comput., 7 (1986), pp. 840–855.
[73] M. Engelman, G. Strang, and K. J. Bathe, The application of quasi-
Newton methods in fluid mechanics, Internat. J. Numer. Methods Engrg., 17 (1981),
pp. 707–718.
[74] V. Faber and T. A. Manteuffel, Necessary and sufficient conditions for the
existence of a conjugate gradient method, SIAM J. Numer. Anal., 21 (1984), pp. 352–
362.
[75] D. Feng, P. D. Frank, and R. B. Schnabel, Local convergence analysis of
tensor methods for nonlinear equations, Math. Programming, 62 (1993), pp. 427–459.
[76] R. Fletcher, Conjugate gradient methods for indefinite systems, in Numerical
Analysis Dundee 1975, G. Watson, ed., Springer-Verlag, Berlin, New York, 1976,
pp. 73–89.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BIBLIOGRAPHY 157
[77] R. W. Freund, A transpose-free quasi-minimal residual algorithm for non-
Hermitian linear systems, SIAM J. Sci. Comput., 14 (1993), pp. 470–482.
[78] R. W. Freund, G. H. Golub, and N. M. Nachtigal, Iterative solution of
linear systems, Acta Numerica, 1 (1991), pp. 57–100.
[79] R. W. Freund, M. H. Gutknecht, and N. M. Nachtigal, An implemen-
tation of the look-ahead Lanczos algorithm for non-Hermitian matrices, SIAM J. Sci.
Comput., 14 (1993), pp. 137–158.
[80] R. W. Freund and N. M. Nachtigal, QMR: a quasi-minimal residual
algorithm for non-hermitian linear systems, Numer. Math., 60 (1991), pp. 315–339.
[81] , An implementation of the QMR method based on coupled two-term
recurrences, SIAM J. Sci. Comput., 15 (1994), pp. 313–337.
[82] D. M. Gay, Some convergence properties of Broyden’s method, SIAM J. Numer.
Anal., 16 (1979), pp. 623–630.
[83] C. W. Gear, Numerical Initial Value Problems in Ordinary Differential Equa-
tions, Prentice-Hall, Englewood Cliffs, NJ, 1971.
[84] R. R. Gerber and F. T. Luk, A generalized Broyden’s method for solving
simultaneous linear equations, SIAM J. Numer. Anal., 18 (1981), pp. 882–890.
[85] P. E. Gill and W. Murray, Quasi-Newton methods for unconstrained
optimization, J. I. M. A., 9 (1972), pp. 91–108.
[86] , Safeguarded steplength algorithms for optimization using descent methods,
Tech. Report NAC 37, National Physical Laboratory Report, Teddington, England,
1974.
[87] P. E. Gill, W. Murray, and M. H. Wright, Practical Optimization,
Academic Press, London, 1981.
[88] A. A. Goldstein, Constructive Real Analysis, Harper and Row, New York,
1967.
[89] G. H. Golub and C. G. VanLoan, Matrix Computations, Johns Hopkins
University Press, Baltimore, 1983.
[90] A. Griewank, Analysis and modification of Newton’s method at singularities,
PhD thesis, Australian National University, 1981.
[91] , Rates of convergence for secant methods on nonlinear problems in Hilbert
space, in Numerical Analysis, Proceedings Guanajuato , Mexico 1984, Lecture Notes
in Math., No, 1230, J. P. Hennart, ed., Springer-Verlag, Heidelberg, 1986, pp. 138–
157.
[92] , The solution of boundary value problems by Broyden based secant methods,
in Computational Techniques and Applications: CTAC 85, Proceedings of CTAC,
Melbourne, August 1985, J. Noye and R. May, eds., North Holland, Amsterdam,
1986, pp. 309–321.
[93] , On the iterative solution of differential and integral equations using secant
updating techniques, in The State of the Art in Numerical Analysis, A. Iserles and
M. J. D. Powell, eds., Clarendon Press, Oxford, 1987, pp. 299–324.
[94] A. Griewank and G. F. Corliss, eds., Automatic Differentiation of Algo-
rithms: Theory, Implementation, and Application, Society for Industrial and Applied
Mathematics, Philadelphia, PA, 1991.
[95] A. Griewank and P. L. Toint, Local convergence analysis for partitioned
quasi-newton updates, Numer. Math., 39 (1982), pp. 429–448.
[96] , Partitioned variable metric methods for large sparse optimization problems,
Numer. Math., 39 (1982), pp. 119–137.
[97] W. A. Gruver and E. Sachs, Algorithmic Methods In Optimal Control,
Pitman, London, 1980.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
158 BIBLIOGRAPHY
[98] M. H. Gutknecht, Variants of BICGSTAB for matrices with complex spectrum,
SIAM J. Sci. Comput., 14 (1993), pp. 1020–1033.
[99] W. Hackbusch, Multi-Grid Methods and Applications, vol. 4 of Springer Series
in Computational Mathematics, Springer-Verlag, New York, 1985.
[100] , Multigrid methods of the second kind, in Multigrid Methods for Integral
and Differential Equations, Oxford University Press, Oxford, 1985.
[101] W. E. Hart and S. O. W. Soul, Quasi-Newton methods for discretized
nonlinear boundary problems, Journal of the Institute of Applied Mathematics, 11
(1973), pp. 351–359.
[102] H. V. Henderson and S. R. Searle, On deriving the inverse of a sum of
matrices, SIAM Rev., 23 (1981), pp. 53–60.
[103] M. R. Hestenes and E. Steifel, Methods of conjugate gradient for solving
linear systems, J. of Res. Nat. Bureau Standards, 49 (1952), pp. 409–436.
[104] D. M. Hwang and C. T. Kelley, Convergence of Broyden’s method in Banach
spaces, SIAM J. Optim., 2 (1992), pp. 505–532.
[105] E. Isaacson and H. B. Keller, Analysis of numerical methods, John Wiley,
New York, 1966.
[106] L. Kantorovich and G. Akilov, Functional Analysis, 2nd ed., Pergamon
Press, New York, 1982.
[107] L. V. Kantorovich, Functional analysis and applied mathematics, Uspehi Mat.
Nauk., 3 (1948), pp. 89–185. translation by C. Benster as Nat. Bur. Standards Report
1509. Washington, D. C., 1952.
[108] H. B. Keller, Newton’s method under mild differentiability conditions, J.
Comput. Sys. Sci, 4 (1970), pp. 15–28.
[109] , Lectures on Numerical Methods in Bifurcation Theory, Tata Institute of
Fundamental Research, Lectures on Mathematics and Physics, Springer-Verlag, New
York, 1987.
[110] C. T. Kelley, Solution of the Chandrasekhar H-equation by Newton’s method,
J. Math. Phys., 21 (1980), pp. 1625–1628.
[111] , A fast multilevel algorithm for integral equations, SIAM J. Numer. Anal.,
32 (1995), pp. 501–513.
[112] C. T. Kelley and E. W. Sachs, A quasi-Newton method for elliptic boundary
value problems, SIAM J. Numer. Anal., 24 (1987), pp. 516–531.
[113] , A pointwise quasi-Newton method for unconstrained optimal control
problems, Numer. Math., 55 (1989), pp. 159–176.
[114] , Fast algorithms for compact fixed point problems with inexact function
evaluations, SIAM J. Sci. Statist. Comput., 12 (1991), pp. 725–742.
[115] , A new proof of superlinear convergence for Broyden’s method in Hilbert
space, SIAM J. Optim., 1 (1991), pp. 146–150.
[116] , Pointwise Broyden methods, SIAM J. Optim., 3 (1993), pp. 423–441.
[117] , Multilevel algorithms for constrained compact fixed point problems, SIAM
J. Sci. Comput., 15 (1994), pp. 645–667.
[118] C. T. Kelley and R. Suresh, A new acceleration method for Newton’s method
at singular points, SIAM J. Numer. Anal., 20 (1983), pp. 1001–1009.
[119] C. T. Kelley and Z. Q. Xue, Inexact Newton methods for singular problems,
Optimization Methods and Software, 2 (1993), pp. 249–267.
[120] , GMRES and integral operators, SIAM J. Sci. Comput., 17 (1996), pp. 217–
226.
[121] T. Kerkhoven and Y. Saad, On acceleration methods for coupled nonlinear
elliptic systems, Numerische Mathematik, 60 (1992), pp. 525–548.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BIBLIOGRAPHY 159
[122] C. Lanczos, Solution of linear equations by minimized iterations, J. Res. Natl.
Bur. Stand., 49 (1952), pp. 33–53.
[123] T. A. Manteuffel, Adaptive procedure for estimating parameters for the
nonsymmetric Tchebychev iteration, Numer. Math., 31 (1978), pp. 183–208.
[124] T. A. Manteuffel and S. Parter, Preconditioning and boundary conditions,
SIAM J. Numer. Anal., 27 (1990), pp. 656–694.
[125] E. S. Marwil, Convergence results for Schubert’s method for solving sparse
nonlinear equations, SIAM J. Numer. Anal., (1979), pp. 588–604.
[126] S. McCormick, Multilevel Adaptive Methods for Partial Differential Equations,
Society for Industrial and Applied Mathematics, Philadelphia, PA, 1989.
[127] J. A. Meijerink and H. A. van der Vorst, An iterative solution method
for linear systems of which the coefficient matrix is a symmetric M-matrix, Math.
Comput., 31 (1977), pp. 148–162.
[128] C. D. Meyer, Matrix Analysis and Applied Linear Algebra, forthcoming.
[129] J. J. Mor´ e, Recent developments in algorithms and software for trust region
methods, in Mathematical Programming: The State of the Art, A. Bachem,
M. Gr¨ oschel, and B. Korte, eds., Springer-Verlag, Berlin, 1983, pp. 258–287.
[130] J. J. Mor´ e and J. A. Trangenstein, On the global convergence of Broyden’s
method, Math. Comput., 30 (1976), pp. 523–540.
[131] T. E. Mott, Newton’s method and multiple roots, Amer. Math. Monthly, 64
(1957), pp. 635–638.
[132] T. W. Mullikin, Some probability distributions for neutron transport in a half
space, J. Appl. Probab., 5 (1968), pp. 357–374.
[133] W. Murray and M. L. Overton, Steplength algorithms for minimizing a
class of nondifferentiable functions, Computing, 23 (1979), pp. 309–331.
[134] N. M. Nachtigal, S. C. Reddy, and L. N. Trefethen, How fast are
nonsymmetric matrix iterations?, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 778–
795.
[135] N. M. Nachtigal, L. Reichel, and L. N. Trefethen, A hybrid gmres
algorithm for nonsymmetric linear systems, SIAM J. Matrix Anal. Appl., 13 (1992).
[136] S. G. Nash, Preconditioning of truncated Newton methods, SIAM J. Sci. Statist.
Comput., 6 (1985), pp. 599–616.
[137] , Who was Raphson? (an answer). Electronic Posting to NA-Digest,
v92n23, 1992.
[138] J. L. Nazareth, Conjugate gradient methods less dependent on conjugacy,
SIAM Rev., 28 (1986), pp. 501–512.
[139] O. Nevanlinna, Convergence of Iterations for Linear Equations, Birkh¨ auser,
Basel, 1993.
[140] I. Newton, The Mathematical Papers of Isaac Newton (7 volumes), D. T.
Whiteside, ed., Cambridge University Press, Cambridge, 1967–1976.
[141] B. Noble, Applied Linear Algebra, Prentice Hall, Englewood Cliffs, NJ, 1969.
[142] J. Nocedal, Theory of algorithms for unconstrained optimization, Acta Numer-
ica, 1 (1991), pp. 199–242.
[143] D. P. O’Leary, Why Broyden’s nonsymmetric method terminates on linear
equations, SIAM J. Optim., 4 (1995), pp. 231–235.
[144] J. M. Ortega, Numerical Analysis A Second Course, Classics in Appl. Math.,
No. 3, Society for Industrial and Applied Mathematics, Philadelphia, PA, 1990.
[145] J. M. Ortega and W. C. Rheinboldt, Iterative Solution of Nonlinear
Equations in Several Variables, Academic Press, New York, 1970.
[146] A. M. Ostrowski, Solution of Equations and Systems of Equations, Academic
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
160 BIBLIOGRAPHY
Press, New York, 1960.
[147] B. N. Parlett, The Symmetric Eigenvalue Problem, Prentice Hall, Englewood
Cliffs, NJ, 1980.
[148] B. N. Parlett, D. R. Taylor, and Z. A. Liu, A look-ahead Lanczos
algorithm for unsymmetric matrices, Math. Comput., 44 (1985), pp. 105–124.
[149] D. W. Peaceman and H. H. Rachford, The numerical solution of parabolic
and elliptic differential equations, J. Soc. Indus. Appl. Math., 11 (1955), pp. 28–41.
[150] G. Peters and J. Wilkinson, Inverse iteration, ill-conditioned equations and
Newton’s method, SIAM Rev., 29 (1979), pp. 339–360.
[151] L. R. Petzold, A description of DASSL: a differential/algebraic system solver,
in Scientific Computing, R. S. Stepleman et al., eds., North Holland, Amsterdam,
1983, pp. 65–68.
[152] E. Picard, M´emoire sur la th´eorie des ´equations aux d´eriv´ees partielles et la
m´ethode des approximations successives, J. de Math. ser 4, 6 (1890), pp. 145–210.
[153] M. J. D. Powell, A hybrid method for nonlinear equations, in Numerical
Methods for Nonlinear Algebraic Equations, Gordon and Breach, New York, 1970,
pp. 87–114.
[154] , Convergence properties of a class of minimization algorithms, in Nonlinear
Programming 2, O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, eds.,
Academic Press, New York, 1975, pp. 1–27.
[155] L. B. Rall, Convergence of the Newton process to multiple solutions, Numer.
Math., 9 (1961), pp. 23–37.
[156] J. Raphson, Analysis aequationum universalis seu ad aequationes algebraicas
resolvendas methodus generalis, et expedita, ex nova infinitarum serierum doctrina,
deducta ac demonstrata. Original in British Library, London, 1690.
[157] G. W. Reddien, On Newton’s method for singular problems, SIAM J. Numer.
Anal., 15 (1978), pp. 993–986.
[158] J. K. Reid, Least squares solution of sparse systems of nonlinear equations by a
modified Marquardt algorithm, in Proc. NATO Conf. at Cambridge, July 1972, North
Holand, Amsterdam, 1973, pp. 437–445.
[159] W. C. Rheinboldt, Numerical Analysis of Parametrized Nonlinear Equations,
John Wiley, New York, 1986.
[160] J. R. Rice, Experiments on Gram-Schmidt orthogonalization, Math. Comput.,
20 (1966), pp. 325–328.
[161] T. J. Rivlin, The Chebyshev Polynomials, John Wiley, New York, 1974.
[162] H. L. Royden, Real Analysis, 2nd ed., Macmillan, New York, 1968.
[163] Y. Saad, Practical use of polynomial preconditionings for the conjugate gradient
method, SIAM J. Sci. Statist. Comput., 6 (1985), pp. 865–881.
[164] , Least squares polynomials in the complex plane and their use for solving
nonsymmetric linear systems, SIAM J. Numer. Anal., 24 (1987), pp. 155–169.
[165] , ILUT: A dual threshold incomplete LU factorization, Tech. Report 92-38,
Computer Science Department, University of Minnesota, 1992.
[166] , A flexible inner-outer preconditioned GMRES algorithm, SIAM J. Sci.
Comput., (1993), pp. 461–469.
[167] Y. Saad and M. Schultz, GMRES a generalized minimal residual algorithm
for solving nonsymmetric linear systems, SIAM J. Sci. Statist. Comput., 7 (1986),
pp. 856–869.
[168] E. Sachs, Broyden’s method in Hilbert space, Math. Programming, 35 (1986),
pp. 71–82.
[169] P. E. Saylor and D. C. Smolarski, Implementation of an adaptive algorithm
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
BIBLIOGRAPHY 161
for Richardson’s method, Linear Algebrra Appl., 154/156 (1991), pp. 615–646.
[170] R. B. Schnabel and P. D. Frank, Tensor methods for nonlinear equations,
SIAM J. Numer. Anal., 21 (1984), pp. 815–843.
[171] E. Schr¨ oder,
¨
Uber unendlich viele Algorithmen zur Auflosung der Gleichungen,
Math. Ann., 2 (1870), pp. 317–365.
[172] L. K. Schubert, Modification of a quasi-Newton method for nonlinear equations
with sparse Jacobian, Math. Comput., 24 (1970), pp. 27–30.
[173] G. A. Schultz, R. B. Schnabel, and R. H. Byrd, A family of trust-region-
based algorithms for unconstrained minimization with strong global convergence
properties, SIAM J. Numer. Anal., 22 (1985), pp. 47–67.
[174] V. E. Shamanskii, A modification of Newton’s method, Ukran. Mat. Zh., 19
(1967), pp. 133–138. (In Russian.)
[175] A. H. Sherman, On Newton-iterative methods for the solution of systems of
nonlinear equations, SIAM J. Numer. Anal., 14 (1978), pp. 755–774.
[176] J. Sherman and W. J. Morrison, Adjustment of an inverse matrix corre-
sponding to changes in the elements of a given column or a given row of the original
matrix (abstract), Ann. Math. Statist., 20 (1949), p. 621.
[177] , Adjustment of an inverse matrix corresponding to a change in one element
of a given matrix, Ann. Math. Statist., 21 (1950), pp. 124–127.
[178] K. Sigmon, MATLAB Primer, Fourth Edition, CRC Press, Boca Raton, FL,
1994.
[179] D. C. Smolarski and P. E. Saylor, An optimum iterative method for solving
any linear system with a square matrix, BIT, 28 (1988), pp. 163–178.
[180] P. Sonneveld, CGS, a fast Lanczos-type solver for nonsymmetric linear
systems, SIAM J. Sci. Statist. Comput., 10 (1989), pp. 36–52.
[181] D. C. Sorensen, Newton’s method with a model trust region modification, SIAM
J. Numer. Anal., 19 (1982), pp. 409–426.
[182] G. Starke, Alternating direction preconditioning for nonsymmetric systems of
linear equations, SIAM J. Sci. Comput., 15 (1994), pp. 369–385.
[183] G. Starke and R. S. Varga, A hybrid Arnoldi-Faber iterative method for
nonsymmetric systems of linear equations, Numer. Math., 64 (1993), pp. 213–239.
[184] G. W. Stewart, Introduction to matrix computations, Academic Press, New
York, 1973.
[185] E. L. Stiefel, Kernel polynomials in linear algebra and their numerical
applications, U.S. National Bureau of Standards, Applied Mathematics Series, 49
(1958), pp. 1–22.
[186] P. N. Swarztrauber, The methods of cyclic reduction, Fourier analysis and
the FACR algorithm for the discrete solution of Poisson’s equation on a rectangle,
SIAM Rev., 19 (1977), pp. 490–501.
[187] , Approximate cyclic reduction for solving Poisson’s equation, SIAM J. Sci.
Statist. Comput., 8 (1987), pp. 199–209.
[188] P. N. Swarztrauber and R. A. Sweet, Algorithm 541: Efficient FORTRAN
subprograms for the solution of elliptic partial differential equations, ACM Trans.
Math. Software, 5 (1979), pp. 352–364.
[189] P. L. Toint, On sparse and symmetric matrix updating subject to a linear
equation, Math. Comput., 31 (1977), pp. 954–961.
[190] J. F. Traub, Iterative Methods for the Solution of Equations, Prentice Hall,
Englewood Cliffs, NJ, 1964.
[191] H. A. van der Vorst, Bi-CGSTAB: A fast and smoothly converging variant to
Bi-CG for the solution of nonsymmetric systems, SIAM J. Sci. Statist. Comput., 13
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
162 BIBLIOGRAPHY
(1992), pp. 631–644.
[192] H. A. van der Vorst and C. Vuik, The superlinear convergence behaviour
of GMRES, Journal Comput. Appl. Math., 48 (1993), pp. 327–341.
[193] R. S. Varga, Matrix Iterative Analysis, Prentice Hall, Englewood Cliffs, NJ,
1962.
[194] E. L. Wachspress, Iterative Solution of Elliptic Systems and Applications to
the Neutron Diffusion Equations of Reactor Physics, Prentice Hall, Englewood Cliffs,
NJ, 1966.
[195] H. F. Walker, Implementation of the GMRES method using Householder
transformations, SIAM J. Sci. Statist. Comput., 9 (1989), pp. 815–825.
[196] , Residual smoothing and peak/plateau behavior in Krylov subspace methods,
Applied Numer. Math., 19 (1995), pp. 279–286.
[197] H. F. Walker and L. Zhou, A simpler GMRES, J. Numer. Lin. Alg. Appl.,
6 (1994), pp. 571–581.
[198] L. T. Watson, S. C. Billups, and A. P. Morgan, Algorithm 652:
HOMPACK: A suite of codes for globally convergent homotopy algorithms, ACM
Trans. Math. Software, 13 (1987), pp. 281–310.
[199] M. A. Woodbury, Inverting modified matrices, Memorandum Report 42,
Statistical Research Group, Princeton NJ, 1950.
[200] D. M. Young, Iterative Solution of Large Linear Systems, Academic Press, New
York, 1971.
[201] T. J. Ypma, The effect of rounding errors on Newton-like methods, IMA J.
Numer. Anal., 3 (1983), pp. 109–118.
[202] L. Zhou and H. F. Walker, Residual smoothing techniques for iterative
methods, SIAM J. Sci. Comput., 15 (1994), pp. 297–312.
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
Index
A-conjugacy, 20
A-norm, 12
Approximate inverse, 6, 77
Armijo rule, 137
Arnoldi process, 37
Backtracking, 136
Bad Broyden method, 132
Banach Lemma, 6
Bi-CG, 47
algorithm, 47
Bi-CGSTAB, 50
algorithm, 50
finite difference, 110
Bi-conjugacy, 47
Bi-Conjugate Gradient, 47
algorithm, 47
Bi-orthogonality, 47
Bounded deterioration, 118, 121
Breakdown, 47
Broyden sequence, 120
Broyden’s method, 113
convergence
linear equations, 118
nonlinear equations, 120
implementation, 123
limited memory, 123
restarting, 123
sparse, 133
Brsol
algorithm, 126
Brsola
algorithm, 145
Cauchy sequence, 67
CGNE, 25
CGNR, 25
CGS, 48
algorithm, 49
Chandrasekhar H-equation, 87
solution
Broyden’s method, 128
Newton’s method, 87
Newton-GMRES, 107
Characteristic polynomial, 34
Chord method, 74
algorithm, 74
convergence, 76
Condition number, 3
Conjugate Gradient
algorithm, 22
finite termination, 14
minimization property, 12
use of residual polynomials, 14
Conjugate Gradient Squared, 48
algorithm, 49
Conjugate transpose, 34
Contraction mapping, 67
Contraction Mapping Theorem, 67
Convection-Diffusion Equation
Broyden’s method, 130
Newton–Armijo, 146
Newton-GMRES, 108
solution
Broyden’s method, 130
Convergence theorem
Broyden’s method
linear equations, 119
163
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
164 INDEX
nonlinear equations, 123
chord method, 76, 77
contraction mapping, 67
finite difference Newton, 81
inexact Newton’s method, 96,
99
Newton’s method, 71
Newton–Armijo, 141
Newton-GMRES, 103
secant method, 82
Shamanskii method, 79
DASPK, 110
Dennis–Mor´e Condition, 114
superlinear convergence, 115
Diagonalizable matrix, 34
Diagonalizing transformation, 34
Difference approximation, 79
choice of h, 80
directional derivative, 80
Jacobian, 80
nonlinearity, 80
Elementary Jordan block, 60
Fixed point, 66
Fixed-point iteration, 66
Forcing term, 95
choice of, 104
Gauss–Seidel
algorithm, 9
Givens rotation, 43
Globally convergent algorithm, 135
GMRES
algorithm, 40, 45
diagonalizable matrices, 34
finite differences, 101
finite termination, 34
GMRES(m), 39, 46
algorithm, 46
loss of orthogonality, 41
minimization property, 33
restarts, 39, 46
use of residual polynomials, 33
Gram–Schmidt process, 37
loss of orthogonality, 40, 41
modified, 40
H-equation, 87
solution
Broyden’s method, 128
Newton’s method, 87
Newton-GMRES, 107
H¨older Continuity, 91
High order methods, 78
Hybrid algorithms, 36
Induced matrix norm
definition, 3
Inexact Newton method, 95
Inner iteration, 100
Inverse power method, 94
Inverse secant equation, 132
Inverse Tangent
Newton–Armijo, 146
Iteration matrix, 5
Iteration statistics, 91
Jacobi
algorithm, 8
convergence result, 8
preconditioning, 9
Jacobi preconditioning, 24
Jordan block, 60
Kantorovich Theorem, 83
Krylov subspace, 11
Left preconditioning, 36
Limited memory, 123
Line Search
parabolic
three-point, 143
two-point, 142
polynomial, 142
Line search, 135, 136
Linear PDE
Bi-CGSTAB, 57
Broyden’s method, 127
CGNR, 55
GMRES, 54
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
INDEX 165
solution
Broyden’s method, 127
TFQMR, 58
Lipschitz continuous
definition, 67
Local improvement, 75
inexact, 100
Matrix norm, 3
Matrix splitting, 7
Modified Bratu Problem, 132
Newton direction, 136
Newton’s method, 71
algorithm, 74
convergence rate, 71
finite differences, 81
stagnation, 81
inexact, 95
monotone convergence, 92
termination, 72
Newton-GMRES, 101
convergence rates, 103
finite differences, 101
safeguarding, 105
Newton-iterative method, 100
Normal matrix, 34
Nsol
algorithm, 86
Nsola
algorithm, 138
Nsolgm
algorithm, 105
Outer iteration, 100
Over solving, 104
Picard iteration, 66
Poisson solver, 24, 27
preconditioner
for CG, 24, 27
for GMRES, 36, 54
Polynomial preconditioning, 24, 36
Richardson iteration, 7
Positive definite matrix, 11
Preconditioned CG
algorithm, 23
Preconditioning
conjugate gradient, 19
GMRES, 36
incomplete
factorization, 24, 36
Jacobi, 24
Poisson solver, 27, 36, 54
polynomial, 24, 36
Richardson iteration, 6
Quasi-minimization, 52
Quasi-Newton methods, 113
Quasi-residual, 51
Quasi-residual norm, 52
Relaxation parameter, 9
Reorthogonalization, 41
Residual
linear, 11
as error estimate, 4
nonlinear
as error estimate, 72
relative nonlinear, 72
Residual polynomials
application to CG, 14
application to CGNR, 25
application to GMRES, 33
application to TFQMR, 52
definition, 14
Richardson iteration, 5
nonlinear, 66
Right preconditioning, 36
Schubert algorithm, 133
Search direction, 136
Secant equation, 113
Secant method, 82
convergence, 82
q-order, 91
r-order, 91
Secant update, 113
Shamanskii method, 78
algorithm, 78
convergence, 79
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.
166 INDEX
finite differences, 91
Sherman–Morrison
formula, 124, 132
Sherman–Morrison–Woodbury for-
mula, 132
Simple decrease, 136
SOR
algorithm, 9
SPD matrix, 12
Spectral radius, 7
Stagnation, 94
Newton’s method, 75
Stationary iterative methods, 5
as preconditioners, 24
Successive substitution, 66
Sufficient decrease, 137
Symmetric Gauss–Seidel
algorithm, 9
preconditioning, 9
Symmetric matrix, 11
Symmetric positive definite matrix,
12
Termination
CG, 16
GMRES, 35, 38
linear equations, 4
nonlinear equations, 72
TFQMR, 52
TFQMR, 51
algorithm, 53
finite difference, 110
weights, 52
Weighted norm, 98
Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

To Polly H. Thomas, 1906-1994, devoted mother and grandmother

1
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

2

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

. . . . . . . . . . . . . . . . . . . . 3 1. .4 Matrix splittings and classical stationary iterative methods . 3. . . . . . . . . . . . . . .7 Examples for preconditioned conjugate iteration 2. 2. . . . .2 CGS. . . . 2. . 3. . . . . . . . . . . . . . . . 3. . . . . . . .1 Krylov methods and the minimization property . . .4 GMRES implementation: Basic ideas . . . . 5 1. . . . .1 The minimization property and its consequences 3. . .6 CGNR and CGNE . . . . . This electronic version is for personal use and may not be duplicated or distributed. CHAPTER 3.1 Review and notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Termination of the iteration . . 3. . . . . . . . . . Conjugate Gradient Iteration 2. . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . .6 Other methods for nonsymmetric systems . . . . . .5 Preconditioning . . . . . .2 Termination .6. .html. . . . . . . . . . . . . . . . . . . . . 2. . . . . . vii Buy this book from SIAM at http://www. .Copyright ©1995 by the Society for Industrial and Applied Mathematics. . . . . . . . . . . . . . . . . . . . 2. .6. . . . . . . . . 7 1. . . . . . . . . . . .2 The Banach Lemma and approximate inverses . . Basic Concepts and Stationary Iterative Methods 3 1. . . 3. . .1 Bi-CG.2 Consequences of the minimization property . . . . . . . . . . . . . . . 7 1. . . .8 Exercises on conjugate gradient . . . . . . GMRES Iteration 3. . . . . . . . 2. . . . . .5 Exercises on stationary iterative methods . . . . . 10 CHAPTER 2. . . . . . . . . . . . . . . . . 2. 3. . . . .com/SIAM/FR16. . . . . . Contents Preface xi How to Get the Software xiii CHAPTER 1. . . . . . .3 The spectral radius . .ec-securehost. . . .5 Implementation: Givens rotations . . . . . . . . . . . . . . . . . . . . . . .6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 11 13 15 19 22 25 26 30 33 33 35 36 37 43 46 47 48 50 . . . . . . . .3 Bi-CGSTAB. . . . . . . . . . . . . . . . .4 Implementation . . . .3 Preconditioning .

. . . . 4. . . .1 Local convergence of Newton’s method 5. . . . . . . . 6. . . . . . . . . . . 6. . . . . . . . . .1. .1 Direct analysis. . . . . . 51 54 55 60 65 65 66 68 71 71 72 73 75 76 77 78 79 82 83 86 91 95 95 95 97 100 100 101 104 104 106 107 108 110 113 114 116 118 120 123 127 CHAPTER 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . Basic Concepts and Fixed-Point Iteration 4. .2. . . . . . . Buy this book from SIAM at http://www. . . Examples for CGNR. . . . 6. CHAPTER 5. . . . . . .2 Newton-iterative methods . . CHAPTER 6. . . . . . . . . . . . .3 The Shamanskii method. . . . . . . . . . . CHAPTER 7. . . . . . .2. . . . . . . 6. . e 7. .1.html. . . . . . . . . . . . .7 3.4 Examples for Newton-GMRES . Examples for GMRES iteration . . . . . . . . . . . . . . .2 Convergence analysis . . . . Inexact Newton Methods 6. . . . . . . . . . . . . .2 Fixed-point iteration . . . . . . . . . .3 Implementation of Newton’s method . . . . . . . . . . . . . . . . . . . . . . . . . . .9 . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Newton GMRES. . . . . . .2 Termination of the iteration .5 The secant method. 6. . . . . . . . . . . .6.2 Convection-diffusion equation. . . . . 7. . . . 5.3 Newton-GMRES implementation . . . . . . 5. . . . 5. . . . . . . . . . . .1 Chandrasekhar H-equation. . .2 Other Newton-iterative methods. . and TFQMR iteration . 6. . . .3 Errors in the function. . . . . . . . . . . . . . . . .1 The basic estimates . . . . . . . 6. . . . .5 The Kantorovich Theorem . .4. . . . . . . . . . . . . 7. . . . . . . . 5. .4 Examples for Broyden’s method . CONTENTS 3. . . . . . . . . . . . . . . . . .1 The chord method. . . 6. . viii 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 TFQMR. . . . .Copyright ©1995 by the Society for Industrial and Applied Mathematics. . . . . . .6 Examples for Newton’s method . . 5. . . . . .4. . . . . . . . . . 6. . . .4. . . . .2 Nonlinear problems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4. . . . . . . . . . . .com/SIAM/FR16. . . . . . . . . .3 The standard assumptions . . . . .4. . . . . . . . . . .7 Exercises on Newton’s method .2 Approximate inversion of F . . .3 Implementation of Broyden’s method 7. . . . . . . . . . . . Newton’s Method 5. . . . . . . . .2. . . . 6. .5 Exercises on inexact Newton methods . . . . . . . . . . . .4 Errors in the function and derivative . . . .1 Linear problems. . . . . . . . . . . . . . . . . Exercises on GMRES . . . . .8 3. . . . Bi-CGSTAB. . . . . . . . . . . . . . . . . .4. . . . . . 5. . . . . . . . . . . . . .1 The Dennis–Mor´ condition . .4 Difference approximation to F .ec-securehost. .1. . . . . . . . . . . . . . . . . . . . . . 5. . This electronic version is for personal use and may not be duplicated or distributed. . . 6. .4. . . . 5. . . . . .4. . . . . . . . . . . . . . . . 7. . . . . . . . . . . . . . Broyden’s method 7. . . . . . . . . 5. .2 Weighted norm analysis. . . . . 5. . . . . . . . . . . . .1 Types of convergence .

4. . . . . . . . . . . . . . . . . . . . . . . . 8.3. CONTENTS ix 7. . . . . . . .ec-securehost. . .4. . . .html. . . .5 7. . . . . . . . . . . . . . . . .2 Analysis of the Armijo rule . .2 Broyden’s method. . . . 8. . . . . . . . . . . . . . . . . .2 Convection-diffusion equation. . . . . . . . 132 . . .1 Polynomial line searches. . . . . 8. . . . . .5 Exercises on global convergence . .4. . . . 8. . . . .com/SIAM/FR16. . . . . . 8. . . .3. .3 Implementation of the Armijo rule .4 Examples for Newton–Armijo . . .1 Single equations . . . . . . . . . . . . . 8. 128 Exercises on Broyden’s method . . . . 8. . . . 8. . . . . . . . . . . . . . . . . . . . . . 135 135 138 141 142 144 146 146 146 148 151 153 163 CHAPTER 8. .1 Inverse tangent function. . . . . . . . . . .Copyright ©1995 by the Society for Industrial and Applied Mathematics. . . . . . . . . . . Bibliography Index Buy this book from SIAM at http://www. .4. . . . . . . 127 7. . . . . . . . . . . .4. . . .1 Linear problems. . . . . . . . . . . . . .3 Broyden–Armijo. This electronic version is for personal use and may not be duplicated or distributed. . . . . . Global Convergence 8. . . . . . . . . . . . . . 8. . . .2 Nonlinear problems. . . . . . . . . . . . . . . . . .

Copyright ©1995 by the Society for Industrial and Applied Mathematics.html. x CONTENTS Buy this book from SIAM at http://www. .ec-securehost.com/SIAM/FR16. This electronic version is for personal use and may not be duplicated or distributed.

the matrix-free formulation and analysis for GMRES and conjugate gradient is almost unchanged in an infinite-dimensional setting.com/SIAM/FR16.html. Inc. The computational examples in this book were done with MATLAB (version 4. [105]. for doing the exercises. We present a limited number of computational examples. These examples are intended to provide results that can be used to validate the reader’s own implementations and to give a sense of how the algorithms perform. or [184].1 on an Apple Macintosh Powerbook 180) and the MATLAB environment is an excellent one for getting experience with the algorithms. A good introduction to the latest 1 MATLAB is a registered trademark of The MathWorks. The analysis of Broyden’s method presented in Chapter 7 and the implementations presented in Chapters 7 and 8 are different from the classical ones and also extend directly to an infinite-dimensional setting. xi Buy this book from SIAM at http://www.ec-securehost. we have selected for coverage mostly algorithms and methods of analysis which extend directly to the infinite-dimensional case and whose convergence can be thoroughly analyzed. . The computational examples and exercises focus on discretizations of infinitedimensional problems such as integral and differential equations. and for small-to-medium scale production work.0a on various SUN SPARCstations and version 4. The examples are not designed to give a complete picture of performance or to be a suite of test problems. It may also be used as a textbook for introductory courses in nonlinear equations or iterative methods or as source material for an introductory course in numerical analysis at the graduate level. linear algebra. This electronic version is for personal use and may not be duplicated or distributed.1 MATLAB codes for many of the algorithms are available by anonymous ftp. For example. We assume that the reader is familiar with elementary numerical analysis. Preface This book on iterative methods for linear and nonlinear equations can be used as a tutorial and a reference by anyone who needs to solve nonlinear systems of equations or large linear systems. Our approach is to focus on a small number of methods and treat them in depth.Copyright ©1995 by the Society for Industrial and Applied Mathematics. and the central ideas of direct methods for the numerical solution of dense linear systems as described in standard texts such as [7]. Though this book is written in a finite-dimensional setting.

1998 Buy this book from SIAM at http://www. Any opinions. I thank Chung-Wei Ng and my parents for over one hundred and ten years of patience and support. Carl Meyer. Zhaqing Xue. Homer Walker. Kelley Raleigh. North Carolina January. Tony Choi. the general algorithmic descriptions or even the MATLAB codes can easily be translated to another language. Jeff Butera. Ilse Ipsen. xii PREFACE version (version 4. C. Belinda King. DMS-9024622 and DMS-9321938. ideas. If the reader has no access to MATLAB or will be solving very large problems. Steve Wright. and an anonymous reviewer for their contributions and encouragement.ec-securehost. Laura Helfrich. and conclusions or recommendations expressed in this material are those of the author and do not necessarily reflect the views of the National Science Foundation or of the Air Force Office of Scientific Research. T. Moody Chu. Casey Miller. This electronic version is for personal use and may not be duplicated or distributed. Ekkehard Sachs.html. Lea Jenkins. I am especially indebted to Jim Banoczi. Mike Tocci. Steve Campbell. suggestions. . Jim Epperson. Yue Zhang. Joseph Skudlarek. Vickie Kearn. and pointers to the literature. Parts of this book are based upon work supported by the National Science Foundation and the Air Force Office of Scientific Research over several years.com/SIAM/FR16. Many of my students and colleagues discussed various aspects of this project with me and provided important corrections. Most importantly. most recently under National Science Foundation Grant Nos. Howard Elman. Andreas Griewank. findings.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Jeff Scroggs. Debbie Lockhart. [43] is also a useful resource. Gordon Wade.2) of MATLAB is the MATLAB Primer [178].

http://www. This electronic version is for personal use and may not be duplicated or distributed.mathworks. from the MathWorks World Wide Web site.mathworks.Copyright ©1995 by the Society for Industrial and Applied Mathematics. .com xiii Buy this book from SIAM at http://www.com WWW: http://www.siam.ec-securehost.html One can obtain MATLAB from The MathWorks.html. MA 01760. The MATLAB codes can be obtained by anonymous ftp from the MathWorks server ftp.com/SIAM/FR16. Phone: (508) 653-1415 Fax: (508) 653-2997 E-mail: info@mathworks.mathworks.com or from SIAM’s World Wide Web site http://www. 24 Prime Park Way Natick. How to get the software A collection of MATLAB codes has been written to accompany this book.org/books/kelley/kelley.com in the directory pub/books/kelley. Inc.

and Induced norms have the important property that Ax ≤ A x .1) Ax = b. x∗ = A−1 b ∈ RN is to be found. Throughout this chapter x will denote a potential solution and {xk }k≥0 the sequence of iterates. b ∈ RN is given.ec-securehost. Recall that the condition number of A relative to the norm κ(A) = A A−1 .1. . Definition 1. x =1 where A is a nonsingular N × N matrix. If  · · is is the lp norm x p = N j=1 1/p |(x)i | p 3 Buy this book from SIAM at http://www. Chapter 1 Basic Concepts and Stationary Iterative Methods 1. Let · be a norm on RN .1. We will write linear equations as (1. The induced matrix norm of an N × N matrix A is defined by A = max Ax . where κ(A) is understood to be infinite if A is singular.com/SIAM/FR16. An excellent reference for the basic ideas of numerical linear algebra and direct methods for linear equations is [184]. Review and notation We begin by setting notation and reviewing some ideas from numerical linear algebra that we expect the reader to be familiar with.1.html. We will denote the ith component of a vector x by (x)i (note the parentheses) and the ith component of xk by (xk )i . In this chapter · will denote a norm on RN as well as the induced matrix norm. This electronic version is for personal use and may not be duplicated or distributed.Copyright ©1995 by the Society for Industrial and Applied Mathematics. We will rarely need to refer to individual components of vectors.

e0 A −1 r0 r0 Ae = A−1 r e r ≤ κ(A) . e0 r0 rk < τ.4) < τ. This electronic version is for personal use and may not be duplicated or distributed. The termination criterion (1.2) which can be related to the error e = x − x∗ in terms of the condition number. x. Let b.3) Proof. 4 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS we will write the condition number as κp . One termination criterion is (1. Buy this book from SIAM at http://www.1. For this reason we prefer to terminate the iteration when rk (1. Since r = b − Ax = −Ae we have e = A−1 Ae ≤ A−1 and r0 = Ae0 ≤ A e0 .2) and (1. r0 as asserted. Lemma 1. .1. (1. particularly when the linear iteration is being used as part of a nonlinear solver. b The two conditions (1.ec-securehost.2) depends on the initial iterate and may result in unnecessary work when the initial iterate is good and a poor result when the initial iterate is far from the solution.4) are the same when x0 = 0.Copyright ©1995 by the Society for Industrial and Applied Mathematics.html. which is a common choice. Hence e A−1 r r ≤ = κ(A) . Let A be nonsingular and let x∗ = A−1 b.com/SIAM/FR16. x0 ∈ RN . Most iterative methods terminate when the residual r = b − Ax is sufficiently small.

In (1.6) xk+1 = (I − A)xk + b. Hence l Ml . Iterative methods of this form are called stationary iterative methods because the transition from xk to xk+1 does not depend on the history of the iteration. l=0 The partial sums Sk = form a Cauchy sequence in RN ×N . Buy this book from SIAM at http://www. k l=0 Ml To see this note that for all m > k m l=k+1 Sk − Sm ≤ Now. Since M Sk + I = Sk+1 . The Banach Lemma and approximate inverses The most straightforward approach to an iterative solution of a linear system is to rewrite (1. One way to do this is to write Ax = b as (1.com/SIAM/FR16. because m l=k+1 · is a matrix norm that is induced by a vector 1 − M m−k 1− M Sk − Sm ≤ M l = M k+1 →0 as m.8) .2. We will discuss more general methods in which {xk } is given by (1. This electronic version is for personal use and may not be duplicated or distributed. Lemma 1. we must have M S + I = S and hence (I − M )S = I.8) holds by showing that the series ∞ M l = (I − M )−1 . All our results are based on the following lemma. 1− M Proof.7) xk+1 = M xk + c. M l ≤ M norm. This proves that I − M is nonsingular and that S = (I − M )−1 . BASIC CONCEPTS 5 1.1) as a linear fixed-point iteration.5) x = (I − A)x + b. . We will show that I − M is nonsingular and that (1. and to define the Richardson iteration (1.7) M is an N ×N matrix called the iteration matrix.ec-securehost. k → ∞.1. The Krylov methods discussed in Chapters 2 and 3 are not stationary iterative methods. If M is an N × N matrix with M < 1 then I − M is nonsingular and 1 (I − M )−1 ≤ (1.2.html.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Hence the sequence Sk converges. say to S.

the matrices B that allow us to apply the Banach lemma and its corollary are called approximate inverses. as Lemma 1. If A and B are N × N matrices and B is an approximate inverse of A.9) and (1. 1 − I − BA B −1 ≤ A .1 I − M = I − (I − BA) = BA is nonsingular. 1 − I − BA A I − BA . It is sometimes possible to precondition a linear equation by multiplying both sides of (1. 1 − I − BA A − B −1 ≤ Proof. If M < 1 then the iteration (1. (1.8) and completes the proof. preconditioned with approximate inversion. This electronic version is for personal use and may not be duplicated or distributed. then not only will the iteration converge rapidly.html.1. then A and B are both nonsingular and (1. Let M = I − BA. By (1. 1 − I − BA B I − BA .1.9). inequality (1.2. By Lemma 1.6) will converge if I − A < 1. Corollary 1.1. Theorem 1. The second part follows in a similar way from B −1 = A(I − M )−1 .1. Hence both A and B are nonsingular.10) A−1 − B ≤ A−1 ≤ B .11) A−1 B −1 = (I − M )−1 ≤ 1 1− M = 1 . 1 − I − BA Since A−1 = (I − M )−1 B. and use (1. .2. A − B −1 = B −1 (I − BA). B is an approximate inverse of A if I − BA < 1.7) converges to x = (I − M )−1 c for all initial iterates x0 . A consequence of Corollary 1.2.com/SIAM/FR16. 6 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Noting that (I − M )−1 ≤ ∞ l=0 M l = (1 − M )−1 . The following theorem is often referred to as the Banach Lemma. but.2.9).ec-securehost.Copyright ©1995 by the Society for Industrial and Applied Mathematics.2. termination decisions based on the Buy this book from SIAM at http://www.1) by a matrix B BAx = Bb so that convergence of iterative methods is improved.1. proves (1. In the context of Richardson iteration.8) (1. Definition 1.2.1 indicates. has the form xk+1 = (I − BA)xk + Bb. Richardson iteration.1 is that Richardson iteration (1.11) implies the first part of (1.12) If the norm of I − BA is small. The following corollary is a direct consequence of Lemma 1. To complete the proof note that A−1 − B = (I − BA)A−1 .

1.7) to the norm of the matrix M .1 is a characterization of convergent stationary iterative methods. This method is a very effective technique for solving differential equations. BASIC CONCEPTS 7 preconditioned residual Bb − BAx will better reflect the actual error. Methods such as Jacobi. Then for any > 0 there is a norm · on RN such that ρ(A) > A − .3.14) has a partial converse that allows us to completely describe the performance of iteration (1. The spectral radius of an N × N matrix A is (1. [100]. The proof is left as an exercise.com/SIAM/FR16.2.3. and sucessive overrelaxation (SOR) iteration are based on splittings of A of the form A = A1 + A2 . Definition 1. We let σ(A) denote the set of eigenvalues of A.2 related convergence of the iteration (1.html.3. The inequality (1. [126].5). Theorem 1. Theorem 1. can also be interpreted in this light. [179]. The term on the right-hand side of the second equality in (1.1. [99].3. and related problems [15]. .1. [111].7) converges for all c ∈ RN if and only if ρ(M ) < 1. [117]. Hence the performance of the iteration is not completely described by M . which tries to approximate A−1 by a polynomial in A [123]. We mention one other approach. Let M be an N ×N matrix.4. Buy this book from SIAM at http://www.1. [169].2. Gauss–Seidel. This electronic version is for personal use and may not be duplicated or distributed. [6]. integral equations. that (1. The iteration (1. in fact. Matrix splittings and classical stationary iterative methods There are ways to convert Ax = b to a linear fixed-point iteration that are different from (1. The concept of spectral radius allows us to make a complete description.14) ρ(A) ≤ A for any induced matrix norm.13) is the limit used by the radical test for convergence of the series An .ec-securehost.13) ρ(A) = max |λ| = lim An λ∈σ(A) n→∞ 1/n .7) in terms of spectral radius.5. and Exercise 1.3. The spectral radius The analysis in § 1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. We state that converse as a theorem and refer to [105] for a proof. polynomial preconditioning. 1.1. Let A be an N × N matrix. It is clear. Lemma 1. A consequence of Theorem 1. Multigrid methods [19]. However the norm of M could be small in some norms and quite large in others. The spectral radius of M is independent of any particular matrix norm of M .

Then Ax = b is converted to the fixedpoint problem x = A−1 (b − A2 x). 8 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS where A1 is a nonsingular matrix constructed so that equations with A1 as coefficient matrix are easy to solve. This electronic version is for personal use and may not be duplicated or distributed. 0< j=i Then A is nonsingular and the Jacobi iteration (1. These methods are usually less efficient than the Krylov methods discussed in Chapters 2 and 3 or the more modern stationary methods based on multigrid ideas.16) |aij | < |aii |.html. or [200].com/SIAM/FR16. Letting (xk )i denote the ith component of the kth iterate we can express Jacobi iteration concretely as   (1. However the classical methods have a role as preconditioners. [144]. where D is the diagonal of A and L and U are the (strict) lower and upper triangular parts.15) converges to x∗ = A−1 b for all b.15) (xk+1 )i = a−1 bi ii − j=i aij (xk )j  .1. [105]. 1 The analysis of the method is based on an estimation of the spectral radius of the iteration matrix M = −A−1 A2 . The limited description in this section is intended as a review that will set some notation to be used later. As a first example we consider the Jacobi iteration that uses the splitting A1 = D. Note that A1 is diagonal and hence trivial to invert. Hence MJAC ∞ < 1 and the iteration converges to the unique solution of x = M x + D−1 b.ec-securehost.Copyright ©1995 by the Society for Industrial and Applied Mathematics. . We present only one convergence result for the classical stationary iterative methods. This leads to the iteration matrix MJAC = −D−1 (L + U ). [193]. Buy this book from SIAM at http://www. Proof. Theorem 1. Also I − M = D−1 A is nonsingular and therefore A is nonsingular. Note that the ith row sum of M = MJAC satisfies N j=1 |mij | = j=i |aij | |aii | < 1.4. Let A be an N × N matrix and assume that for all 1≤i≤N (1. 1 For a detailed description of the classical stationary iterative methods the reader may consult [89]. A2 = L + U.

[193]. one wants to write the stationary method as a preconditioned Richardson iteration. The performance can be dramatically improved with a good choice of ω but still is not competitive with Krylov methods. For the Jacobi iteration. A symmetric Gauss–Seidel iteration is a forward Gauss–Seidel iteration followed by a backward Gauss–Seidel iteration. [8].com/SIAM/FR16. That means that one wants to find B such that M = I − BA and then use B as an approximate inverse.18) BSGS = (D + LT )−1 D(D + L)−1 . If A is symmetric then U = LT . A1 = D + L. Note that A1 is lower triangular. BASIC CONCEPTS 9 Gauss–Seidel iteration overwrites the approximate solution with the new value as soon as it is computed. A2 = U. From the point of view of preconditioning.html.ec-securehost. A further disadvantage is that the choice of ω is often difficult to make. and iteration matrix MBGS = −(D + U )−1 L. [89].17) For symmetric Gauss–Seidel (1. and hence A−1 y is easy to compute for vectors 1 y. (1. and the papers cited therein provide additional reading on this topic. MGS = −(D + L)−1 U. In that event MSGS = (D + U )−1 L(D + L)−1 U = (D + LT )−1 L(D + L)−1 LT . References [200]. The successive overrelaxation iteration modifies Gauss–Seidel by adding a relaxation parameter ω to construct an iteration with iteration matrix MSOR = (D + ωL)−1 ((1 − ω)D − ωU ). resulting in the splitting A1 = D + U. BJAC = D−1 . This electronic version is for personal use and may not be duplicated or distributed. unlike Jacobi iteration. Buy this book from SIAM at http://www. Backward Gauss–Seidel begins the update of x with the N th coordinate rather than the first. A2 = L. This leads to the iteration matrix MSGS = MBGS MGS = (D + U )−1 L(D + L)−1 U.Copyright ©1995 by the Society for Industrial and Applied Mathematics. . the iteration depends on the ordering of the unknowns. This results in the iteration   (xk+1 )i = a−1 bi − ii the splitting and iteration matrix aij (xk+1 )j − j<i j>i aij (xk )j  . Note also that.

5. Buy this book from SIAM at http://www. This electronic version is for personal use and may not be duplicated or distributed. Verify equality (1. Exercises on stationary iterative methods 1.ec-securehost. 10 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 1. Show that if A is symmetric and positive definite (that is AT = A and xT Ax > 0 for all x = 0) that BSGS is also symmetric and positive definite.com/SIAM/FR16.5.2.5.2. Prove Theorem 1. 1. .5.4.3.18).7) fails to converge. Show that if ρ(M ) ≥ 1 then there are x0 and c such that the iteration (1.3.Copyright ©1995 by the Society for Industrial and Applied Mathematics.1. 1. 1.html.5.

com/SIAM/FR16. . Recall that A is symmetric if A = AT and positive definite if xT Ax > 0 for all x = 0. The two such methods that we’ll discuss in depth. Brief descriptions of several of these methods and their properties are in § 3. some measure of error over the affine space x0 + Kk . There are other Krylov methods that are not as well understood as CG or GMRES. [12].6.ec-securehost. . Krylov methods and the minimization property In the following two chapters we describe some of the Krylov space methods for linear equations. So {rk }k≥0 will denote the sequence of residuals rk = b − Axk . .1.html. The residual is r = b − Ax. It has come into wide use over the last 15 years as an iterative method and has generally superseded the Jacobi–Gauss–Seidel–SOR family of methods. we assume that A is a nonsingular N × N matrix and let x∗ = A−1 b. and [78]. Unlike the stationary iterative methods. where x0 is the initial iterate and the kth Krylov subspace Kk is Kk = span(r0 . The conjugate gradient (CG) iteration was invented in the 1950s [103] as a direct method. . This electronic version is for personal use and may not be duplicated or distributed. Ak−1 r0 ) for k ≥ 1. Chapter 2 Conjugate Gradient Iteration 2.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Ar0 . minimize. at the kth iteration. 11 Buy this book from SIAM at http://www. . conjugate gradient and GMRES. As in Chapter 1. CG is intended to solve symmetric positive definite (spd) systems. Krylov methods do not have an iteration matrix.

2) 1 φ(x) = xT Ax − xT b 2 over x0 + Kk . preconditioning. and at the very end. Lemma 2. A If e = x − x∗ then e 2 A = eT Ae = (A(x − x∗ ))T A−1 (A(x − x∗ )) = b − Ax 2 A−1 and hence the A-norm of the error is also the A−1 -norm of the residual. we begin with a description of what the algorithm does and the consequences of the minimization property of the iterates. Since (x∗ )T Ax∗ is independent of x. · A is called the A-norm. Buy this book from SIAM at http://www. ˜ Minimizing φ over any subset of RN is the same as minimizing x − x∗ A over that subset. Since A is spd we may define a norm (you should check that this is a norm) by (2. If xk minimizes φ over S then xk also minimizes x∗ − x A = r A−1 over S. performance.1. In this section. Note that if φ(˜) is the minimal value (in RN ) then x ∇φ(˜) = A˜ − b = 0 x x and hence x = x∗ . The development in these notes is different from the classical work and more like the analysis for GMRES and CGNR in [134]. We will use this lemma in the particular case that S = x0 + Kk for some k. Since A is symmetric and Ax∗ = b −xT Ax∗ − (x∗ )T Ax = −2xT Ax∗ = −2xT b.1. .ec-securehost. The kth iterate xk of CG minimizes (2.1) x A = √ xT Ax. Therefore x − x∗ 2 A = 2φ(x) + (x∗ )T Ax∗ . This electronic version is for personal use and may not be duplicated or distributed. After that we describe termination criterion.Copyright ©1995 by the Society for Industrial and Applied Mathematics.html. Note that x − x∗ 2 A = (x − x∗ )T A(x − x∗ ) = xT Ax − xT Ax∗ − (x∗ )T Ax + (x∗ )T Ax∗ . Let S ⊂ RN . We state this as a lemma. Proof. minimizing φ is equivalent to minimizing x − x∗ 2 and hence to minimizing x − x∗ A .com/SIAM/FR16. and in the section on GMRES that follows. 12 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS In this section we assume that A is spd. the implementation.

Since U U T = U T U = I by orthogonality of U . we have Aj = U Λj U T .3) x∗ − xk A ≤ x∗ − w A for all w ∈ x0 + Kk . 2 A Define A1/2 = U Λ1/2 U T and note that (2.2.4) x∗ − xk A = p∈Pk . Hence (2.Copyright ©1995 by the Society for Industrial and Applied Mathematics.4) Pk denotes the set of polynomials of degree k.p(0)=1 min p(A)(x∗ − x0 ) A.5) x = xT Ax = A1/2 x 2 .1 implies that since xk minimizes φ over x0 + Kk (2. Since any w ∈ x0 + Kk can be written as w= k−1 j=0 γj Aj r0 + x0 for some coefficients {γj }.com/SIAM/FR16. 2 Buy this book from SIAM at http://www. we can express x∗ − w as x∗ − w = x∗ − x0 − Since Ax∗ = b we have r0 = b − Ax0 = A(x∗ − x0 ) and therefore x∗ − w = x∗ − x0 − where the polynomial p(z) = 1 − k−1 j=0 k−1 j=0 k−1 j=0 γj Aj r0 . This electronic version is for personal use and may not be duplicated or distributed. where U is an orthogonal matrix whose columns are the eigenvectors of A and Λ is a diagonal matrix with the positive eigenvalues of A on the diagonal.ec-securehost. CONJUGATE GRADIENT ITERATION 13 2. γj z j+1 has degree k and satisfies p(0) = 1. The spectral theorem for spd matrices asserts that A = U ΛU T . Consequences of the minimization property Lemma 2. γj Aj+1 (x∗ − x0 ) = p(A)(x∗ − x0 ). Hence p(A) = U p(Λ)U T . .1.html. In (2.

let i=1 p(z) = ¯ N i=1 (λi − z)/λi . Proof. When doing that we seek to terminate the iteration when some specified error tolerance is reached. for any x ∈ RN and p(A)x A = A1/2 p(A)x 2 ≤ p(A) 2 A1/2 x 2 ≤ p(A) 2 x A. It is best to regard CG as an iterative method.6) xk − x∗ A ≤ x0 − x∗ A p∈Pk .2. . In that way we showed (in the absence of all roundoff error!) that CG terminated in finitely many iterations with the exact solution. Theorem 2. One simple application of (2. Corollary 2. This leads to an upper estimate for the number of CG iterations required to reduce the A-norm of the error to a given tolerance.Copyright ©1995 by the Society for Industrial and Applied Mathematics. since in most applications the number of unknowns N is very large. Let {λi }N be the eigenvalues of A. by (2.1.7) and the fact ¯ that p vanishes on σ(A). Here σ(A) is the set of all eigenvalues of A.ec-securehost. This electronic version is for personal use and may not be duplicated or distributed. Hence. This. Then the CG algorithm will find the solution within N iterations.4) implies that (2. The following corollary is an important consequence of (2. based on information on σ(A). 14 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Hence. ¯ xN − x∗ A ≤ x0 − x∗ A z∈σ(A) max |¯(z)| = 0. Let A be spd and let {xk } be the CG iterates.2. ¯ ¯ p ∈ PN because p has degree N and p(0) = 1. Let A be spd. Then p ¯ (2. This is not as good as it sounds.html. Let k be given and let {¯k } be any kth degree polynomial such that pk (0) = 1.2.1. p z∈σ(A) We will refer to the polynomial pk as a residual polynomial [185].1. together with (2.com/SIAM/FR16.} In specific contexts we try to construct sequences of residual polynomials.p(0)=1 z∈σ(A) min max |p(z)|.8) Pk = {p | p is a polynomial of degree k and p(0) = 1.7) easy to evaluate. and one cannot afford to perform N iterations. ¯ Definition 2.7) xk − x∗ x0 − x∗ A A ≤ max |¯k (z)|.6). The set of kth degree residual polynomials is (2. p Note that our test polynomial had the eigenvalues of A as its roots.7) is to show how the CG algorithm can be viewed as a direct method. Buy this book from SIAM at http://www. As a test polynomial. that make either the middle or the right term in (2.

Then the CG iteration for Ax = b with x0 = 0 will terminate in at most k iterations. Proof.4) and the fact that x0 = 0 that xk − x∗ A ≤ p(A)x∗ ¯ A = 0. but rather terminate once some criterion has been satisfied.2. . By the spectral theorem l=1 x∗ = k l=1 (γl /λil )uil .7). If the spectrum of A has fewer than N points. Assume that there are exactly k ≤ N distinct eigenvalues of A.9) b − Axk 2 ≤ η b 2 . we have by (2. Let A be spd. Termination of the iteration In practice we do not run the CG iteration until an exact solution is found.3.ec-securehost.2. This electronic version is for personal use and may not be duplicated or distributed. we can use a similar technique to prove the following theorem. Let A be spd with eigenvectors {ui }N . p(λil ) = 0 for 1 ≤ l ≤ k and ¯ hence p(A)x∗ = ¯ k l=1 p(λil )γl /λil uil = 0. ¯ One can easily verify that p ∈ Pk . Theorem 2. We use the residual polynomial. The error estimates that come from the minimization property. Moreover. p(z) = ¯ k l=1 (λil − z)/λil . Theorem 2. Buy this book from SIAM at http://www. are based on (2. Let b be a linear i=1 combination of k of the eigenvectors of A b= k l=1 γl uil . This completes the proof. Then the CG iteration terminates in at most k iterations. CONJUGATE GRADIENT ITERATION 15 In the two examples that follow we look at some other easy consequences of (2. 2.2. however. ¯ So. One typical criterion is small (say ≤ η) relative residuals.com/SIAM/FR16.3. This means that we terminate the iteration after (2.Copyright ©1995 by the Society for Industrial and Applied Mathematics.html.7) and therefore estimate the reduction in the relative A-norm of the error. Let {λil } be the eigenvalues of A associated with the eigenvectors {uil }k .

. b − Axk b − Ax0 2 2 = A(x∗ − xk ) A(x∗ − x0 ) 2 2 ≤ λ1 x∗ − xk λN x∗ − x0 A A as asserted.html.1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. . λN > 0. to predict the performance of the CG iteration based on termination on small relative residuals. We will do this in the next two lemmas and then illustrate the point with an example.13) follows directly from (2. i = λN ≤ Az ≤ λ1 N T 2 i=1 λi (ui z) 2 2 = N 2 T 2 i=1 λi (ui z) N T 2 i=1 λi (ui z) = λ1 A1/2 z 2 . first recall that if A = U ΛU T is the spectral decomposition of A and we order the eigenvalues such that λ1 ≥ λ2 ≥ . Let ui be a unit eigenvector corresponding to λi .11) twice.10) and (2. This electronic version is for personal use and may not be duplicated or distributed. (2.com/SIAM/FR16. = z T Az = (A1/2 z)T (A1/2 z) = A1/2 z 2 2 which proves (2. we must not only use (2.13) b r0 2 2 b − Axk b 2 b − Axk b 2 2 = b − Axk b − Ax0 2 2 ≤ κ2 (A) xk − x∗ x∗ − x0 xk − x∗ x∗ − x0 A A A A 2 ≤ κ2 (A) r0 b 2 2 . using (2. So κ2 (A) = λ1 /λN .3.7) to predict when the relative Buy this book from SIAM at http://www. The equality on the left of (2. So. Lemma 2. . Then for all z ∈ RN . Clearly z 2 A λN 1/2 z A ≤ Az 2 ≤ λ1 1/2 z A.10) complete the proof. A1/2 z 2 = z A (2.11) Proof.12).10) and (2.2.ec-securehost. 16 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Our next task is to relate the relative residual in the Euclidean norm to the relative error in the A-norm. λN . Let A be spd with eigenvalues λ1 ≥ λ2 ≥ . Therefore.12) and (2. then A 2 = λ1 and A−1 2 = 1/λN . 2 Taking square roots and using (2.3.12) is clear and (2. To obtain the inequality on the right of (2.12). . We may write A = U ΛU T as Az = Hence λN A1/2 z 2 2 N i=1 λi (uT z)ui . .10). Lemma 2. Proof.

after k iterations (2. ≤√ κ2 (A) + 1 11 + 3 Buy this book from SIAM at http://www. √ √ Hence. If we let pk (z) = ¯ (10 − z)k /10k . after k iterations we have Axk − b 2 √ ≤ 11 × 10−k /3.Copyright ©1995 by the Society for Industrial and Applied Mathematics. CONJUGATE GRADIENT ITERATION 17 A-norm error is small. Assume that x0 = 0 and that the eigenvalues of A are contained in the interval (9. We consider a very simple example. A result from [48] (see also [45]) that is.2 we simply note that κ2 (A) ≤ 11/9. but also use Lemma 2. So.3. ∞).05. This means that we may apply (2. Note that it is always the case that the spectrum of a spd matrix is contained in the interval [λN . then pk ∈ Pk . Hence. √ κ2 (A) − 1 11 − 3 ≈ . when k ≥ 3.7) to get ¯ xk − x∗ It is easy to see that Hence. that is. p 9≤z≤11 max |¯k (z)| = 10−k .7).2 to relate small A-norm errors to small relative residuals. the sharpest possible. in one sense. the size of the A-norm of the error will be reduced by a factor of 10−3 when 10−k ≤ 10−3 .14) xk − x∗ A A ≤ x∗ A 9≤z≤11 max |¯k (z)|. p ≤ x∗ A 10 −k . 11). . when k ≥ 4.html. that is. is (2. λ1 ] and that κ2 (A) = λ1 /λN .3. To use Lemma 2. One can obtain a more precise estimate by using a polynomial other than pk in the upper estimate for the right-hand side of (2. b 2 So.15) xk − x ∗ A ≤ 2 x0 − x ∗ A κ2 (A) − 1 κ2 (A) + 1 k . we can estimate κ2 (A) by κ2 (A) ≤ 11/9. since ( x − 1)/( x + 1) is an increasing function of x on the interval (1. In the case of the above example.ec-securehost.com/SIAM/FR16. the size of the relative residual will be reduced by a factor of 10−3 when √ 10−k ≤ 3 × 10−3 / 11. This electronic version is for personal use and may not be duplicated or distributed.

p which is a sharper estimate on convergence.15) can be very pessimistic. the estimate in (2.3 ≈ 2. Exercise 2.91) ≈ 3.3.25/1.9). which. A ≤ 10−3 x∗ A. From the results above one can see that if the condition number of A is near one.3/1. the condition number can be quite large.2)k .2)k ≤ 10−3 or when k > −3/ log10 (. We will illustrate this with an example.15) gives xk − x∗ x∗ A A ≤ 2 × (19/21)k ≈ 2 × (. Hence (2. In fact.com/SIAM/FR16.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 18 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Therefore (2. 1.5.04 = 82. which also predicts termination within three iterations. Based on this information the best estimate of the condition number of A is κ2 (A) ≤ 400. when inserted into (2. if we use as a residual polynomial p3k ∈ P3k ¯ p3k (z) = ¯ It is easy to see that z∈σ(A) (1. (1.html. .5 also covers this point.25 − z)k (400 − z)2k . We may have more precise information than a single interval containing σ(A).15) predicts 83 iterations and the clustering analysis 15 (the smallest integer multiple of 3 larger than 3 × 4.91)k < 10−3 or when k > − log10 (2000)/ log10 (. 400).15) would predict that xk − x∗ when 2 × (.6.05) ≈ 3.25)k = (.05k < 10−3 or when k > − log10 (2000)/ log10 (. the CG iteration will converge very rapidly. If the eigenvalues cluster in a small number of intervals.3 = 12. This electronic version is for personal use and may not be duplicated or distributed.3/. However.ec-securehost. (2. Assume that x0 = 0 and the eigenvalues of A lie in the two intervals (1.8.15) would predict a reduction in the size of the A-norm error by a factor of 10−3 when 2 × .2) = 4. The estimate based on the clustering gives convergence in 3k iterations when (. When we do.25)k × 4002k max |¯3k (z)| ≤ (.5) and (399.91)k . but CG can perform very well. This would indicate fairly slow convergence. Even if the condition number Buy this book from SIAM at http://www.

1. unless the l iteration terminates prematurely.17) αk+1 = pT (b − Axk ) p T rk k+1 = T k+1 .18).16) can be written as pT Axk + αpT Apk+1 − pT b = 0 k+1 k+1 k+1 leading to (2. .ec-securehost. Once pk+1 has been found.1).1 then implies that T T rk 2 = rk rk = rk rk+1 = 0 and hence xk = x∗ . Now.e. easier to solve) is called preconditioning. Let A be spd and let {xk } be the conjugate gradient iterates. the iteration will perform well if the eigenvalues are clustered in a few small intervals. We used this term before in the context of Richardson iteration and accomplished the goal by multiplying A by an approximate inverse. then rk = rk+1 . αk+1 is easy to compute from the minimization property of the iteration.4. Recalling that ∇φ(x) = Ax − b = −r we have (2. for any ξ ∈ Kk dφ(xk + tξ) = ∇φ(xk + tξ)T ξ = 0 dt at t = 0. pT Apk+1 pk+1 Apk+1 k+1 If xk = xk+1 then the above analysis implies that α = 0. proves that (if we define p0 = 0) pT rk = 0 for all 0 ≤ l < k ≤ n. CONJUGATE GRADIENT ITERATION 19 is large. We discuss this in § 2. if xk = xk+1 . We show that this only happens if xk is the solution. (2. such a simple approach can destroy the symmetry of the coefficient matrix and a more subtle implementation is required.19) T ∇φ(xk )T ξ = −rk ξ = 0 for all ξ ∈ Kk .5. In fact dφ(xk + αpk+1 ) (2. this proves (2.8. 2 The next lemma characterizes the search direction and. Buy this book from SIAM at http://www. In the context of CG. Since rl ∈ Kk for all l < k (see Exercise 2. This electronic version is for personal use and may not be duplicated or distributed.html. Lemma 2. as a side effect.4. 2. Lemma 2. we have.4. either xk = x∗ or a search direction pk+1 = 0 can be found very cheaply so that xk+1 = xk + αk+1 pk+1 for some scalar αk+1 ..18) Proof. The transformation of the problem into one with eigenvalues clustered near one (i.com/SIAM/FR16. Equation (2. Then T rk rl = 0 for all 0 ≤ l < k. Since xk minimizes φ on x0 + Kk .16) =0 dα for the correct choice of α = αk+1 . Implementation The implementation of CG depends on the amazing fact that once xk has been determined.Copyright ©1995 by the Society for Industrial and Applied Mathematics.

20 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Lemma 2.Copyright ©1995 by the Society for Industrial and Applied Mathematics.4.22) αk+1 pT Aξ = −(Axk − b)T ξ = −∇φ(xk )T ξ = 0. any k+1 pk+1 satisfying (2.e.23) Proof. By Lemma 2. Define p0 = 0. be expressed as pk+1 = rk + wk with wk ∈ Kk .ec-securehost. Lemma 2.2 then implies that pT Arl = 0 for 0 ≤ l ≤ k − 2. k+1 This uniquely specifies the direction of pk+1 as (2. then rl ∈ Kl+1 ⊂ Kk−1 .24) T βk+1 = −rk Ark−1 /pT Ark−1 k provided pT Ark−1 = 0. Trivially k+1 (2. Since Kk ⊂ Kk+1 . in the scalar product (x.4. .1. Theorem 2. (2. rk−1 ). Let A be spd and assume that rk = 0. (2. (2. k+1 It only remains to solve for βk+1 so that pT Ark−1 = 0. . Let pk+1 be given by (2. a subspace of dimension one less than Kk+1 .20) pk+1 ∈ Kk+1 .2. up to a scalar multiple. Now.23). y) = xT Ay) to Kk . k+1 k If l ≤ k − 2. pT Aξ = 0 for all ξ ∈ Kk .2 and the fact that Kk = span(r0 . We have the following theorem.23) then pT Arl = 0 k+1 for all 0 ≤ l ≤ k − 1. While one might think that wk would be hard to compute. trivial. Let A be spd and let {xk } be the conjugate gradient iterates..21) then imply that for all ξ ∈ Kk .21) ∇φ(xk+1 )T ξ = (Axk + αk+1 Apk+1 − b)T ξ = 0 for all ξ ∈ Kk .20) can.4. . it is. .com/SIAM/FR16. in fact. we need only verify that a βk+1 can be found so that if pk+1 is given by (2.19) and (2. k+1 Proof. The condition pT Aξ = 0 is called A-conjugacy of pk+1 to Kk .4. Since k rk = rk−1 − αk Apk Buy this book from SIAM at http://www. This electronic version is for personal use and may not be duplicated or distributed. (2. Then for any l ≤ k T pT Arl = rk Arl + βk+1 pT Arl .22) implies that pk+1 ∈ Kk+1 is A-orthogonal (i. If xk = x∗ then xk+1 = xk + αk+1 pk+1 and pk+1 is determined up to a scalar multiple by the conditions (2. Then pk+1 = rk + βk+1 pk for some βk+1 and k ≥ 0.html. .

31) pT Apk = pT A(rk−1 + βk pk−1 ) k k = pT Ark−1 + βk pT Apk−1 = pT Ark−1 . Let A be spd and assume that rk = 0. To get (2.com/SIAM/FR16.27) note that pT Apk = 0 and hence (2. Note that for k ≥ 0 (2.30) Also note that (2.4. k k k βk+1 = T −rk Apk . k 2 This completes the proof. rk−1 2 2 Buy this book from SIAM at http://www.30). pT Apk k Now combine (2.ec-securehost.26).28) is that pT rk = 0 and k hence pT rk = (rk + βk+1 pk )T rk = rk 2 . Lemma 2.29) gives T 0 = pT rk − αk+1 pT Apk+1 = rk k+1 k+1 2 2 − αk+1 pT Apk+1 .html.23) implies that k+1 (2. CONJUGATE GRADIENT ITERATION 21 2 2 we have T rk rk−1 = rk−1 − αk pT Ark−1 .1 we have (2. This electronic version is for personal use and may not be duplicated or distributed. .28) T pT rk+1 = rk rk+1 + βk+1 pT rk+1 = 0 k+1 k αk+1 = rk 2 2 pT Apk+1 k+1 rk rk−1 2 2 2.4.29) k+1 2 Taking scalar products of both sides of rk+1 = rk − αk+1 Apk+1 with pk+1 and using (2. (2.4. Then (2.27) Proof.2. An immediate consequence of (2. (2. The common implementation of conjugate gradient uses a different form for αk and βk than given in (2.25) to get βk+1 = T −rk Apk αk . and (2. k T Since rk rk−1 = 0 by Lemma 2.24).17) and (2. k+1 which is equivalent to (2. 2 βk+1 = by Lemma 2.3.Copyright ©1995 by the Society for Industrial and Applied Mathematics.25) pT Ark−1 = rk−1 2 /αk = 0.31).26) and (2.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

22

ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS

Now take scalar products of both sides of rk = rk−1 − αk Apk with rk and use Lemma 2.4.1 to get rk
2 2 T = −αk rk Apk .

Hence (2.27) holds. The usual implementation reflects all of the above results. The goal is to find, for a given , a vector x so that b−Ax 2 ≤ b 2 . The input is the initial iterate x, which is overwritten with the solution, the right hand side b, and a routine which computes the action of A on a vector. We limit the number of iterations to kmax and return the solution, which overwrites the initial iterate x and the residual norm. Algorithm 2.4.1. cg(x, b, A, , kmax) 1. r = b − Ax, ρ0 = r 2 , k = 1. 2 √ 2. Do While ρk−1 > b 2 and k < kmax (a) if k = 1 then p = r else β = ρk−1 /ρk−2 and p = r + βp (b) w = Ap (c) α = ρk−1 /pT w (d) x = x + αp (e) r = r − αw (f) ρk = r
2 2

(g) k = k + 1 Note that the matrix A itself need not be formed or stored, only a routine for matrix-vector products is required. Krylov space methods are often called matrix-free for that reason. Now, consider the costs. We need store only the four vectors x, w, p, and r. Each iteration requires a single matrix-vector product (to compute w = Ap), two scalar products (one for pT w and one to compute ρk = r 2 ), and three 2 operations of the form ax + y, where x and y are vectors and a is a scalar. It is remarkable that the iteration can progress without storing a basis for the entire Krylov subspace. As we will see in the section on GMRES, this is not the case in general. The spd structure buys quite a lot. 2.5. Preconditioning To reduce the condition number, and hence improve the performance of the iteration, one might try to replace Ax = b by another spd system with the same solution. If M is a spd matrix that is close to A−1 , then the eigenvalues

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

CONJUGATE GRADIENT ITERATION

23

of M A will be clustered near one. However M A is unlikely to be spd, and hence CG cannot be applied to the system M Ax = M b. In theory one avoids this difficulty by expressing the preconditioned problem in terms of B, where B is spd, A = B 2 , and by using a two-sided preconditioner, S ≈ B −1 (so M = S 2 ). Then the matrix SAS is spd and its eigenvalues are clustered near one. Moreover the preconditioned system SASy = Sb has y ∗ = S −1 x∗ as a solution, where Ax∗ = b. Hence x∗ can be recovered from y ∗ by multiplication by S. One might think, therefore, that computing S (or a subroutine for its action on a vector) would be necessary and that a matrix-vector multiply by SAS would incur a cost of one multiplication by A and two by S. Fortunately, this is not the case. If y k , rk , pk are the iterate, residual, and search direction for CG applied ˆ ˆ to SAS and we let ˆ ˆ ˆ ˆ xk = S y k , rk = S −1 rk , pk = S pk , and zk = S rk , then one can perform the iteration directly in terms of xk , A, and M . The reader should verify that the following algorithm does exactly that. The input is the same as that for Algorithm cg and the routine to compute the action of the preconditioner on a vector. Aside from the preconditioner, the arguments to pcg are the same as those to Algorithm cg. Algorithm 2.5.1. pcg(x, b, A, M, , kmax) 1. r = b − Ax, ρ0 = r 2 , k = 1 2 √ 2. Do While ρk−1 > b 2 and k < kmax (a) z = M r (b) τk−1 = z T r (c) if k = 1 then β = 0 and p = z else β = τk−1 /τk−2 , p = z + βp (d) w = Ap (e) α = τk−1 /pT w (f) x = x + αp (g) r = r − αw (h) ρk = rT r (i) k = k + 1 Note that the cost is identical to CG with the addition of • the application of the preconditioner M in step 2a and • the additional inner product required to compute τk in step 2b.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

24

ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS

Of these costs, the application of the preconditioner is usually the larger. In the remainder of this section we briefly mention some classes of preconditioners. A more complete and detailed discussion of preconditioners is in [8] and a concise survey with many pointers to the literature is in [12]. Some effective preconditioners are based on deep insight into the structure of the problem. See [124] for an example in the context of partial differential equations, where it is shown that certain discretized second-order elliptic problems on simple geometries can be very well preconditioned with fast Poisson solvers [99], [188], and [187]. Similar performance can be obtained from multigrid [99], domain decomposition, [38], [39], [40], and alternating direction preconditioners [8], [149], [193], [194]. We use a Poisson solver preconditioner in the examples in § 2.7 and § 3.7 as well as for nonlinear problems in § 6.4.2 and § 8.4.2. One commonly used and easily implemented preconditioner is Jacobi preconditioning, where M is the inverse of the diagonal part of A. One can also use other preconditioners based on the classical stationary iterative methods, such as the symmetric Gauss–Seidel preconditioner (1.18). For applications to partial differential equations, these preconditioners may be somewhat useful, but should not be expected to have dramatic effects. Another approach is to apply a sparse Cholesky factorization to the matrix A (thereby giving up a fully matrix-free formulation) and discarding small elements of the factors and/or allowing only a fixed amount of storage for the factors. Such preconditioners are called incomplete factorization preconditioners. So if A = LLT + E, where E is small, the preconditioner is (LLT )−1 and its action on a vector is done by two sparse triangular solves. We refer the reader to [8], [127], and [44] for more detail. One could also attempt to estimate the spectrum of A, find a polynomial p such that 1 − zp(z) is small on the approximate spectrum, and use p(A) as a preconditioner. This is called polynomial preconditioning. The preconditioned system is p(A)Ax = p(A)b and we would expect the spectrum of p(A)A to be more clustered near z = 1 than that of A. If an interval containing the spectrum can be found, the residual polynomial q(z) = 1 − zp(z) of smallest L∞ norm on that interval can be expressed in terms of Chebyshev [161] polynomials. Alternatively q can be selected to solve a least squares minimization problem [5], [163]. The preconditioning p can be directly recovered from q and convergence rate estimates made. This technique is used to prove the estimate (2.15), for example. The cost of such a preconditioner, if a polynomial of degree K is used, is K matrix-vector products for each application of the preconditioner [5]. The performance gains can be very significant and the implementation is matrix-free.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

for any residual polynomial pk ∈ Pk .html.32) asserts that x∗ − x 2 AT A = (x∗ − x)T AT A(x∗ − x) = (Ax∗ − Ax)T (Ax∗ − Ax) = (b − Ax)T (b − Ax) = r 2 is minimized over x0 + Kk at each iterate. ¯ (2.32) AT Ax = AT b. The third. Conjugate Gradient on the Normal equations to minimize the Error. This approach [103] is called CGNR [71]. The reason for this name is that the minimization property of CG as applied to (2.33) asserts that if y ∗ is the solution to (2. when we consider the AT A norm of the error. disadvantage is that one must compute the action of AT on a vector as part of the matrix-vector product involving AT A. one could solve (2. There are three disadvantages that may or may not be serious.33) then y∗ − y 2 AAT = (y ∗ − y)T (AAT )(y ∗ − y) = (AT y ∗ − AT y)T (AT y ∗ − AT y) = x∗ − x 2 2 is minimized over y0 + Kk at each iterate.34) rk 2 ≤ pk (AT A)r0 ¯ 2 ≤ r0 2 z∈σ(AT A) max |¯k (z)|. [134]. p Buy this book from SIAM at http://www. 2 Hence. The advantages of this approach are that all the theory for CG carries over and the simple implementation for both CG and PCG can be used. The reason for this name is that the minimization property of CG as applied to (2.33) AAT y = b and then set x = AT y to solve Ax = b. CONJUGATE GRADIENT ITERATION 25 2. As we will see in the chapter on nonlinear problems. one might consider solving Ax = b by applying CG to the normal equations (2. . This approach [46] is now called CGNE [78]. CGNR and CGNE If A is nonsingular and nonsymmetric. more important.com/SIAM/FR16.ec-securehost. As above. We consider the case for CGNR. the analysis for CGNE is essentially the same. The second is that two matrix-vector products are needed for each CG iterate since w = AT Ap = AT (Ap) in CGNR and w = AAT p = A(AT p) in CGNE.6. we have x∗ − x 2 AT A = (x∗ − x)T AT A(x∗ − x) = A(x∗ − x) 2 2 = r 2. Hence the name Conjugate Gradient on the Normal equations to minimize the Residual. The analysis with residual polynomials is similar to that for CG. [78]. Alternatively. [134]. there are situations where this is not possible. The first is that the condition number of the coefficient matrix AT A is the square of that of A.Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

MATLAB functions for the matrix-vector product and (optionally) the preconditioner. We set u0j = u(n+1)j = ui0 = ui(n+1) = 0.34) and (2. On return the code supplies the approximate solution x and the history of the iteration as the vector of residual norms. are the initial iterate.7. Hence the performance of CGNR and CGNE is determined by the distribution of singular values. The unknowns are uij ≈ u(xi . Examples for preconditioned conjugate iteration In the collection of MATLAB codes we provide a code for preconditioned conjugate gradient iteration. y) = 0.html.ec-securehost.36) = (αij + α(i+1)j )(u(i+1)j − uij ) −(α(i−1)j + αij )(uij − u(i−1)j ) + (αi(j+1) + αij )(ui(j+1) − uij ) −(αij + αi(j−1) )(uij − ui(j−1) ) Buy this book from SIAM at http://www. xj ) where xi = ih for 1 ≤ i ≤ n. which corresponds exactly to the termination criterion. the residual polynomial is to be maximized over the eigenvalues of AT A. and define αij = −a(xi . the right hand side vector b. . One can verify [105] that the differential operator is positive definite in the Hilbert space sense and that the five-point discretization described below is positive definite if a > 0 for all 0 ≤ x. y)∇u) = f (x. Most significantly. The inputs.7).com/SIAM/FR16. y) on 0 < x.3. The estimate is in terms of the l2 norm of the residual. y) = u(1. We discretize with a five-point centered difference scheme with n2 points and mesh width h = 1/(n + 1). We consider the discretization of the partial differential equation (2. which is the set of the squares of the singular values of A. 1) = u(0. 0) = u(x. 2. 0 < x. to reflect the boundary conditions. hence we need not prove a result like Lemma 2. y ≤ 1 (Exercise 2. We express the discrete matrix-vector product as (Au)ij (2.10).35) ˙ −∇(a(x.2.Copyright ©1995 by the Society for Industrial and Applied Mathematics. xj )h−2 /2. y < 1 subject to homogeneous Dirichlet boundary conditions u(x.8. y < 1. x0 . described in the comment lines. and iteration parameters to specify the maximum number of iterations and the termination criterion. This electronic version is for personal use and may not be duplicated or distributed. 26 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS There are two major differences between (2.

This electronic version is for personal use and may not be duplicated or distributed. For a preconditioner we use a Poisson solver. and the representation as a one-dimensional array. We allowed up to 100 CG iterations. For the MATLAB implementation we convert freely from the representation of u as a two-dimensional array (with the boundary conditions added). We will report our results graphically.html. subject to homogeneous Dirichlet boundary conditions. The effectiveness of such a preconditioner has been analyzed in [124] and some of the many ways to implement the solver efficiently are discussed in [99]. See the routine fish2d in collection of MATLAB codes for an example of how to do this in MATLAB. and [187]. This example illustrates the importance of a good preconditioner. The properties of CG on the preconditioned problem in the continuous case have been analyzed in [48]. j ≤ n.15) indicates that convergence will be slow. The initial iterate was the zero vector. which is what pcgsol expects to see. designed to take advantage of fast MATLAB built-in functions.9. is more efficient that the classical stationary iterative methods.5 ). This can be explained by the fact that the eigenvalues are not clustered and κ(A) = O(1/h2 ) = O(n2 ) = O(N ) and hence (2. in the case of the MATLAB environment used in this book. [186]. The initial iterate was u0 = 0. In Fig. Note that the reduction in r is not monotone. In the results reported here we took n = 31 resulting in a system with N = n2 = 961 unknowns. [188].1 the solid line is a plot of rk / b and the dashed line a plot of u∗ −uk A / u∗ −u0 A . plotting rk 2 / b 2 on a semi-log scale. For the computations reported in this section we took a(x. The reader is asked to quantify this in terms of execution times in Exercise 2. CONJUGATE GRADIENT ITERATION 27 for 1 ≤ i.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Because of this their execution time is less than a simple count of floatingpoint operations would indicate. however. 2.com/SIAM/FR16. which predicts decrease in e A but not necessarily in r as the iteration progresses. Poisson solvers can be designed to take advantage of vector and/or parallel architectures or. . For many types of domains and boundary conditions.8.ec-securehost. The fast Poisson solver in the collection of Buy this book from SIAM at http://www. By this we mean an operator G such that v = Gw is the solution of the discrete form of −vxx − vyy = w. Note that the unpreconditioned iteration is slowly convergent. This is consistent with the theory. Even the unpreconditioned iteration. which is useful for computing the action of A on u and applying fast solvers. y) = cos(x) and took the right hand side so that the exact solution was the discretization of 10xy(1 − x)(1 − y) exp(x4. We expect second-order accuracy from the method and accordingly we set termination parameter = h2 = 1/1024.

PCG for 2-D elliptic equation. 28 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 1 10 Relative Residual and A-norm of Error 10 0 10 -1 10 -2 10 -3 10 -4 0 10 20 30 Iterations 40 50 60 Fig.1.2. The preconditioned iteration required 5 iterations for convergence and the unpreconditioned iteration 52.com/SIAM/FR16.html. 2. This electronic version is for personal use and may not be duplicated or distributed. CG for 2-D elliptic equation. the built-in function fft. 2.2 the solid line is the graph of rk 2 / b 2 for the preconditioned iteration and the dashed line for the unpreconditioned. . In Fig.ec-securehost. codes fish2d is based on the MATLAB fast Fourier transform. We caution the reader that the preconditioned iteration is not as much faster than the Buy this book from SIAM at http://www. Not only does the preconditioned iteration converge more rapidly. 2. 10 1 10 0 Relative Residual 10 -1 10 -2 10 -3 10 -4 0 10 20 30 Iterations 40 50 60 Fig. but the number of iterations required to reduce the relative residual by a given amount is independent of the mesh spacing [124].Copyright ©1995 by the Society for Industrial and Applied Mathematics.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. CONJUGATE GRADIENT ITERATION 29 unpreconditioned one as the iteration count would suggest. the execution time of the preconditioned iteration was about 60% of that of the unpreconditioned.ec-securehost. . In Exercise 2.2 million floating-point operations while the preconditioned iteration required .11 you are asked to compare execution times for your own environment.html. Hence. In the MATLAB environment we used.8. the cost of the preconditioner is considerable.87 million floating-point operations. As we remarked above. Buy this book from SIAM at http://www. The MATLAB flops command indicates that the unpreconditioned iteration required roughly 1. This electronic version is for personal use and may not be duplicated or distributed.com/SIAM/FR16. this speed is a result of the efficiency of the MATLAB fast Fourier transform.

y) > 0 for all 0 ≤ x. . Exercises on conjugate gradient 2. 2.8.8.1) ∪ (2.5. 2. Compare the execution times on your computing environment (using the cputime command in MATLAB.com/SIAM/FR16. 30 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 2. for instance). What happens as N is increased? How are the performance √ and accuracy affected by changes in a(x.8. If κ(A) = O(N ).8. 2.3.7 for example. Let A be spd. Buy this book from SIAM at http://www. 2. Show that if A has constant diagonal then PCG with Jacobi preconditioning produces the same iterates as CG with no preconditioning. Show that there is a spd B such that B 2 = A. Give upper estimates based on (2.11. 2. in MATLAB by writing the matrix-vector product routines and using the MATLAB codes pcgsol and fish2d.6) for the number of CG iterations required to reduce the A norm of the error by a factor of 10−3 and for the number of CG iterations required to reduce the residual by a factor of 10−3 . 2. Is B unique? 2. Prove that rl ∈ Kk for all l < k.10. Duplicate the results in § 2.2). Explain your findings. 2).8.8.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 2. Let Λ be a diagonal matrix with Λii = λi and let p be a polynomial.7.html. 2. give a rough estimate of the number of CG iterates required to reduce the relative residual to O(1/N ). Let {xk } be the conjugate gradient iterates. This electronic version is for personal use and may not be duplicated or distributed.8. assume that the cost of a matrix vector multiply is 4N floating-point multiplies. nonsingular. Prove Theorem 2.1 + x and examine the accuracy of the result. 1. Assume that A is spd and that σ(A) ⊂ (1. Let A be a nonsingular matrix with all singular values in the interval (1. Estimate the number of CGNR/CGNE iterations required to reduce the relative residual by a factor of 10−4 . 2. Prove that the linear transformation given by (2.6.36) is symmetric and 2 positive definite on Rn if a(x.9. y) = .8.1.8.3. y ≤ 1.ec-securehost.8.8. For the matrix A in problem 5.8. Estimate the number of floatingpoint operations reduce the A norm of the error by a factor of 10−3 using CG iteration.2. Prove that p(Λ) = maxi |p(λi )| where · is any induced matrix norm. and spd. 2.8. y)? Try a(x.2.4. Assume that A is N × N .

0 < x. Plot φ(xn ) as a function of n for each of the examples.Copyright ©1995 by the Society for Industrial and Applied Mathematics.15. Implement Jacobi (1. This electronic version is for personal use and may not be duplicated or distributed. Jacobi. y) − uyy (x. < 1. subject to the inhomogeneous Dirichlet boundary conditions u(x. How does the performance compare to CG and PCG? 2.12. y) + ex+y u(x. 2. y < 1. y) = 1. Apply CG and PCG to solve the five-point discretization of −uxx (x.14.7. y) = 0. Modify pcgsol so that φ(x) is computed and stored at each iterate and returned on output.7.17) and symmetric Gauss–Seidel (1.ec-securehost. 2. and symmetric Gauss–Seidel).8. y) = 1.8.8.com/SIAM/FR16. CONJUGATE GRADIENT ITERATION 31 2.13. Buy this book from SIAM at http://www. Use the Jacobi and symmetric Gauss–Seidel iterations from Chapter 1 to solve the elliptic boundary value problem in § 2. y. 0) = u(x. u(0. 0 < x. 1) = u(1. Compare the performance with respect to both computer time and number of iterations to preconditioning with the Poisson solver.html. Experiment with different mesh sizes and preconditioners (Fast Poisson solver. .18) preconditioners for the elliptic boundary value problem in § 2.8.

This electronic version is for personal use and may not be duplicated or distributed.html.com/SIAM/FR16. 32 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Buy this book from SIAM at http://www. .Copyright ©1995 by the Society for Industrial and Applied Mathematics.ec-securehost.

com/SIAM/FR16. The kth (k ≥ 1) iteration of GMRES is the solution to the least squares problem minimizex∈x0 +Kk b − Ax 2 . Chapter 3 GMRES Iteration 3. This electronic version is for personal use and may not be duplicated or distributed.1. Moreover. ¯ From this we have the following corollary. GMRES does not require computation of the action of AT on a vector.1) The beginning of this section is much like the analysis for CG. Theorem 3. Note that if x ∈ x0 + Kk then x = x0 + and so b − Ax = b − Ax0 − k−1 j=0 k−1 j=0 k j=1 γj Aj r0 γj A j+1 r0 = r0 − γj−1 Aj r0 . Corollary 3.ec-securehost. This is a significant advantage in many cases.Copyright ©1995 by the Society for Industrial and Applied Mathematics.1. We have just proved the following result. The minimization property and its consequences The GMRES (Generalized Minimum RESidual) was proposed in 1986 in [167] as a Krylov subspace method for nonsymmetric systems.3) rk r0 2 2 ≤ pk (A) 2 . ¯ 33 Buy this book from SIAM at http://www. The use of residual polynomials is made more complicated because we cannot use the spectral theorem to decompose A.2) rk 2 = min p(A)r0 ¯ p∈Pk 2 ≤ pk (A)r0 2 . Let A be nonsingular and let xk be the kth GMRES iteration.1. Unlike CGNR.html. Let A be nonsingular and let xk be the kth GMRES iteration.1. . (3.1. ¯ ¯ Hence if x ∈ x0 + Kk then r = p(A)r0 where p ∈ Pk is a residual polynomial. and therefore storage requirements increase as the iteration progresses. one must store a basis for Kk . Then for all pk ∈ Pk ¯ (3. Then for all pk ∈ Pk ¯ (3.

1. Hence we will switch to complex matrices and vectors. p has degree N . Proof.Copyright ©1995 by the Society for Industrial and Applied Mathematics. By ¯ (3. In that case the columns of V are the eigenvectors of A and V −1 = V H . In Chapter 2 we applied the spectral theorem to obtain more precise information on convergence rates. Recall that the scalar product in C N .4) rk r0 2 2 ≤ κ2 (V ) max |¯k (z)|. This electronic version is for personal use and may not be duplicated or distributed. Let pk ∈ Pk . Let A be nonsingular. The characteristic polynomial of A is p(z) = det(A − zI). Theorem 3. Let A = V ΛV −1 be a nonsingular diagonalizable matrix. rN = b − AxN = 0 and hence xN is the solution. 2 by ≤ V 2 V −1 2 pk (Λ) ¯ 2 ≤ κ2 (V ) max |¯k (z)|. Recall that A is diagonalizable if there is a nonsingular (possibly complex!) matrix V such that A = V ΛV −1 . In the remainder of this section we must use complex arithmetic to analyze the convergence.1.2.3. Here V H is the complex conjugate transpose of V . We pay a price in that we must use complex arithmetic for the only time in this book. Let xk be the kth GMRES iterate.2) to get information from clustering of the spectrum just like we did for CG. Theorem 3. However. . if A is diagonalizable we may use (3. the space of complex N -vectors. we will use the l2 norm in C N . Our use of complex arithmetic will be implicit for the most part and is needed only so that we may admit the possibility of complex eigenvalues of A.com/SIAM/FR16. If A is a diagonalizable matrix and p is a polynomial then p(A) = V p(Λ)V −1 A is normal if the diagonalizing transformation V is orthogonal. In particular. 34 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS We can apply the corollary to prove finite termination of the GMRES iteration. and so pN (z) = p(z)/p(0) ∈ PN ¯ is a residual polynomial. Then for all pk ∈ Pk ¯ (3. p(0) = det(A) = 0 since A is nonsingular.ec-securehost. Then the GMRES algorithm will find the solution within N iterations.html. This is not an option for general nonsymmetric matrices.3). p z∈σ(A) 2 ¯ Proof. We can use the structure of a diagonalizable matrix to prove the following result. We can easily estimate pk (A) ¯ pk (A) ¯ as asserted. p z∈σ(A) Buy this book from SIAM at http://www. It is well known [141] that p(A) = pN (A) = 0. Here Λ is a (possibly complex!) diagonal matrix with the eigenvalues of A on the diagonal. is xH y.

It is a direct ¯ rk 2 ≤ ρk r0 2 .6) illustrates the potential benefits of a good approximate inverse preconditioner.2. 11). GMRES ITERATION 35 It is not clear how one should estimate the condition number of the diagonalizing transformation if it exists. we look at some easy consequences of (3. We assume that x0 = 0 and hence r0 = b.4. Buy this book from SIAM at http://www. the chances are very high that a computed matrix will be diagonalizable. Assume that A has only k distinct eigenvalues.3. The convergence rate estimates for the diagonalizable case will involve κ2 (V ). we terminate the iteration when (3. but will otherwise resemble those for CG. GMRES is best thought of as an iterative method.5.6) 2 ≤ ρ < 1.4). rate estimates have been derived in [139].3) and (3. If A is normal. Using the residual polynomial pk (z) = (10 − z)k /10k we find ¯ rk r0 2 2 ≤ (100)10−k = 102−k . Again we look at examples. The estimate (3.1. This is particularly so for the finite difference Jacobian matrices we consider in Chapters 6 and 8.1. Theorem 3. [134]. As the set of nondiagonalizable matrices has measure zero in the space of N × N matrices. Then GMRES will terminate in at most k iterations.2) that (3. Assume that A = V ΛV −1 is diagonalizable.html. Hence we confine our attention to diagonalizable matrices. We can use (3. If A is not diagonalizable. Let A be a nonsingular diagonalizable matrix.5) holds when 102−k < η or when k > 2 + log10 (η). Theorem 3.ec-securehost. [33]. that the eigenvalues of A lie in the interval (9. [192]. Then the GMRES iteration. As we did for CG. Hence (3. Let pk (z) = (1 − z)k . and [34].4) directly to estimate the first k such that (3.com/SIAM/FR16. Assume that I − A consequence of (3.5) rk 2 ≤η b 2 for the purposes of this example. of course. for Ax = b will terminate in at most k iterations. As was the case with CG. and that κ2 (V ) = 100. 3.3) and (3.2.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Termination As is the case with CG.5) holds without requiring a lemma like Lemma 2. with x0 = 0. κ2 (V ) = 1. Let b be a linear combination of k of the eigenvectors of A b= k l=1 γl uil . This electronic version is for personal use and may not be duplicated or distributed. Let A be a nonsingular normal matrix. .

Hence. [34]. if M A has a smaller condition number than A. If r = M Ax − M b is the residual for the preconditioned system. we have. Again we mention [8] and [12] as a good general references for preconditioning. The residual of the preconditioned problem is the same as that of the unpreconditioned problem. [33]. or GMRES in the normal case. Right preconditioning has been used as the basis for a method that changes the preconditioner as the iteration progresses [166]. Preconditioning Preconditioning for GMRES and other methods for nonsymmetric problems is different from that for CG. [164].ec-securehost. Some hybrid algorithms use the GMRES/Arnoldi process itself to construct polynomial preconditioners for GMRES or for Richardson iteration [135]. and [36] for discussions of and pointers to additional references to several questions related to nonnormal matrices.1. There is no concern that the preconditioned system be spd and hence (3. unlike PCG.7 we use the Poisson solver preconditioner for nonsymmetric partial differential equations. CGNE. If one can find M such that I − MA 2 = ρ < 1. Multigrid [99] and alternating direction [8]. Buy this book from SIAM at http://www. One important aspect of implementation is that. [120]. one can solve the system AM y = b with GMRES and then set x = M y. CGNR. one can apply the algorithm directly to the system M Ax = M b (or AM y = b). .1. if the product M A can be formed without error. This electronic version is for personal use and may not be duplicated or distributed. However there are two different ways to view preconditioning. the value of the relative residuals as estimators of the relative error is unchanged.3. If I − AM 2 = ρ < 1. Preconditioning done in this way is called left preconditioning. 3. Most of the preconditioning ideas mentioned in § 2. we might expect the relative residual of the preconditioned system to be a better indicator of the relative error than the relative residual of the original system.com/SIAM/FR16.Copyright ©1995 by the Society for Industrial and Applied Mathematics. [182] methods have similar performance and may be more generally applicable.5 are useful for GMRES as well. Incomplete factorization (LU in this case) preconditioners can be used [165] as can polynomial preconditioners. then applying GMRES to M Ax = M b allows one to apply (3. This is a very active area of research and we refer to [134].html. by Lemma 1.6) essentially tells the whole story.6) to the preconditioned system. one can write a single matrix-vector product routine for M A (or AM ) that includes both the application of A to a vector and that of the preconditioner. 36 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS The convergence estimates for GMRES in the nonnormal case are much less satisfying that those for CG. [72]. This is called right preconditioning. In the examples in § 3. [183]. Hence. Hence. ek e0 2 2 ≤ κ2 (M A) rk r0 2 2 .

The data are vectors x0 and b. For i = 1. The algorithm computes an orthonormal basis for Kk and stores it in the columns of V . Then any z ∈ Kk can be written as z= vlk k l=1 yl vlk . . A.1. Hence. V ) 1. a map that computes the action of A on a vector. however. is the lth column of Vk . This electronic version is for personal use and may not be duplicated or distributed. Define r0 = b − Ax0 and v1 = r0 / r0 2 . Algorithm 3. Suppose one had an orthogonal projector Vk onto Kk . 2. say. This is a standard linear least squares problem that could be solved by a QR factorization.html. The Gram–Schmidt procedure for formation of an orthonormal basis for Kk is called the Arnoldi [4] process. GMRES implementation: Basic ideas Recall that the least squares problem defining the kth GMRES iterate is minimizex∈x0 +Kk b − Ax 2 . k − 1 vi+1 = Avi − Avi − i T j=1 ((Avi ) vj )vj i T j=1 ((Avi ) vj )vj 2 If there is never a division by zero in step 2 of Algorithm arnoldi. GMRES ITERATION 37 3. . one can represent (3. Buy this book from SIAM at http://www.4.7) minimizey∈Rk r0 − AVk y 2 . If one uses Gram–Schmidt orthogonalization. and a dimension k.1) to a least squares where problem in Rk for the coefficient vector y of z = x − x0 . b.4. then the columns of the matrix Vk are an orthonormal basis for Kk . Hence we can convert (3. arnoldi(x0 . Since x − x0 = Vk y for some y ∈ Rk . The problem with such a direct approach is that the matrix vector product of A with the basis matrix Vk must be taken at each iteration.Copyright ©1995 by the Society for Industrial and Applied Mathematics.ec-securehost. we must have xk = x0 + Vk y where y minimizes b − A(x0 + Vk y) 2 = r0 − AVk y 2 . k.com/SIAM/FR16. our least squares problem in Rk is (3. . . . A division by zero is referred to as breakdown and happens only if the solution to Ax = b is in x0 + Kk−1 .7) very efficiently and the resulting least squares problem requires no extra multiplications of A with vectors.

≤ b 2. Since the columns of Vi are an orthonormal basis for Ki .. hij = 0 if i > j + 1. This electronic version is for personal use and may not be duplicated or distributed.4. . Since xi − x0 ∈ Ki there is y ∈ Ri such that xi − x0 = Vi y. where y k minimizes βe1 − Hk y 2 over Rk .com/SIAM/FR16. (3. Hence xk = x0 + Vk y k . using the orthogonality of Vk+1 .8) Hence. and let i be the smallest integer for which Avi − i j=1 ((Avi )T vj )vj = 0. .html. H is nonsingular since A is. Buy this book from SIAM at http://www. which overwrites the initial iterate x and the residual norm. where H is an i × i matrix. rk = b − Axk = r0 − A(xk − x0 ) = Vk+1 (βe1 − Hk y k ). for a given . the k +1×k matrix Hk whose entries are hij is upper Hessenberg. If the Arnoldi process does not break down. for some y k ∈ Rk . Since r0 = βVi e1 and Vi is an orthogonal matrix ri 2 = Vi (βe1 − Hy) 2 = βe1 − Hy Ri+1 . we can use it to implement GMRES in an efficient way. the right-hand side b. Let A be nonsingular. 38 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Lemma 3. We limit the number of iterations to kmax and return the solution. a vector x so that b − Ax 2 AVk = Vk+1 Hk . 0. where · Rk+1 denotes the Euclidean norm in Rk+1 . x. By hypothesis Avi ∈ Ki and hence AKi ⊂ Ki . rk 2 = Vk+1 (βe1 − Hk y k ) 2 = βe1 − Hk y k Rk+1 . Setting β = r0 and e1 = (1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. The input is the initial iterate. By the Gram–Schmidt construction. . and a map that computes the action of A on a vector. Then x = A−1 b ∈ x0 + Ki . The upper Hessenberg structure can be exploited to make the solution of the least squares problems very efficient [167].e. we have AVi = Vi H. The Arnoldi process (unless it terminates prematurely with a solution) produces matrices {Vk } with orthonormal columns such that (3.1. Proof.ec-securehost. . Set hji = (Avj )T vi . Setting y = βH −1 e1 proves that ri = 0 by the minimization property. Note that when y k has been computed. 0)T ∈ Ri we have ri 2 2 = b − Axi 2 = r0 − A(xi − x0 ) 2 . We have. the norm of rk can be found without explicitly forming xk and computing rk = b − Axk . i. .9) The goal of the iteration is to find. let the vectors vj be generated by Algorithm arnoldi.

This means that in order to perform k GMRES iterations one must store k vectors of length N . There are more efficient ways to solve the least squares problem in step 2f. Note that xk is only computed upon termination and is not needed within the iteration. k = 0 2. . (g) ρ = βe1 − Hk y k 3. ρ) 1. A. b. 0)T (f) e1 = ∈ Rk+1 Minimize βe1 − Hk y k Rk+1 over Rk to obtain y k . when it works it can significantly reduce the storage costs of the iteration. GMRES ITERATION 39 Algorithm 3.com/SIAM/FR16. There is no general convergence theorem for the restarted algorithm and restarting will slow the convergence down. .html. . . xk = x0 + Vk y k . [197]. (3. Buy this book from SIAM at http://www.Copyright ©1995 by the Society for Industrial and Applied Mathematics.k = vk+1 k j=1 hjk vj 2 2 (e) vk+1 = vk+1 / vk+1 (1. A partial remedy is to replace the classical Gram–Schmidt orthogonalization in Algorithm gmresa with modified Gram–Schmidt orthogonalization. will not hold and the residual and approximate solution could be inaccurate. If this happens. call GMRES again with x0 = xk . k hjk = (Avk )T vj (c) vk+1 = Avk − (d) hk+1. Step 2f can be done with a single MATLAB command. While ρ > b 2 and k < kmax do (a) k = k + 1 (b) for j = 1.2. For very large problems this becomes prohibitive and the iteration is restarted when the available room for basis vectors is exhausted. [167]. It is an important property of GMRES that the basis for the Krylov space must be stored as the iteration progress. which is often the case for large problems. which depends on this orthogonality. call GMRES and explicitly test the residual b−Axk if k = m upon termination. We replace Rk+1 . β = ρ. This restarted version of the algorithm is called GMRES(m) in [167]. However. v1 = r/ r 2 . This electronic version is for personal use and may not be duplicated or distributed. 0. One way to implement this is to set kmax to the maximum number m of vectors that one can store.4. at a cost of O(k 3 ) floating-point operations. the result from the previous call. kmax. . Algorithm gmresa can be implemented very straightforwardly in MATLAB.9). We discuss implementation of GMRES(m) later in this section. the backward division operator. . . If the norm of the residual is larger than . gmresa(x. . and we use the method of [167] in the collection of MATLAB codes.ec-securehost. and the simple one-line MATLAB approach can be efficient for such problems. r = b − Ax. . ρ = r 2 . The savings are slight if k is small relative to N . A more serious problem with the implementation proposed in Algorithm gmresa is that the vectors vj may become nonorthogonal as a result of cancellation errors. .

0436e − 07   −1. ρ) 1.0436e − 07 9. j. .0436e − 07 1. One can take a second pass through the modified Gram–Schmidt process and restore lost orthogonality [147]. The outline of our implementation is Algorithm gmresb. This electronic version is for personal use and may not be duplicated or distributed. δ 0 δ We orthogonalize the columns of A with classical Gram–Schmidt to obtain 1. 40 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS the loop in step 2c of Algorithm gmresa with vk+1 = Avk for j = 1. While modified Gram–Schmidt and classical Gram–Schmidt are equivalent in infinite precision. k = 0 2. First.0000e + 00  . . For modified Gram–Schmidt     1. ρ = r 2 . This implementation solves the upper Hessenberg least squares problem using the MATLAB backward division operator.0000e − 07 −1. Second. and is not particularly efficient. . The versions we implement in the collection of MATLAB codes use modified Gram–Schmidt. this version is very simple and illustrates some important ideas. r = b − Ax.com/SIAM/FR16.3.0456e − 14 1.0000e − 07 1. We illustrate this point with a simple example from [128].0456e − 14 1.0000e + 00 4.ec-securehost. . there is an opportunity to compensate for a loss of orthogonality in the basis vectors for the Krylov space.html. b. the modified form is much more likely in practice to maintain orthogonality of the basis. doing the computations in MATLAB. v1 = r/ r 2 . [160]. V =  1. β = ρ. we see that xk need only be computed after termination as the least squares residual ρ can be used to approximate the norm of the residual (they are identical in exact arithmetic).0000e + 00 1. A.9905e − 01  . k T vk+1 = vk+1 − (vk+1 vj )vj . We present a better implementation in Algorithm gmres.3565e − 16 T Here |vi vj − δij | ≤ 10−8 for all i. .   Algorithm 3.0000e − 07 1.3568e − 02 T The columns of VU are not orthogonal at all.9715e − 08   −9.Copyright ©1995 by the Society for Industrial and Applied Mathematics. While ρ > b 2 and k < kmax do (a) k = k + 1 Buy this book from SIAM at http://www. V =  1. However.0000e + 00 1. kmax.004.0000e + 00 4.4. gmresb(x. Let δ = 10−7 and define 1 1 1   A =  δ δ 0 . In fact v2 v3 ≈ −.0000e − 07 −1.

.com/SIAM/FR16. k T i. . one can augment the modified Gram–Schmidt process • vk+1 = Avk for j = 1. . GMRES ITERATION 41 (b) vk+1 = Avk for j = 1. This electronic version is for personal use and may not be duplicated or distributed. . . Even if modified Gram–Schmidt orthogonalization is used. k T hjk = vk+1 vj vk+1 = vk+1 − hjk vj • hk+1. . k T hjk = vk+1 vj vk+1 = vk+1 − hjk vj • hk+1. 0)T (e) e1 = ∈ Rk+1 Minimize βe1 − Hk y k Rk+1 to obtain y k ∈ Rk .k = vk+1 2 Rk+1 . One can test for loss of orthogonality [22].k = vk+1 2 • If loss of orthogonality is detected For j = 1. [147]. . (f) ρ = βe1 − Hk y k 3.k = vk+1 2 2 (d) vk+1 = vk+1 / vk+1 (1. The modified Gram–Schmidt process with reorthogonalization looks like • vk+1 = Avk for j = 1. . There is nothing to be gained by reorthogonalizing more than once [147]. . . .Copyright ©1995 by the Society for Industrial and Applied Mathematics. xk = x0 + Vk y k . . . • vk+1 = vk+1 / vk+1 2 with a second pass (reorthogonalization).k = vk+1 2 2 • vk+1 = vk+1 / vk+1 Buy this book from SIAM at http://www. and reorthogonalize if needed or use a more stable means to create the matrix V [195]. These more complex implementations are necessary if A is ill conditioned or many iterations will be taken. . k T htmp = vk+1 vj hjk = hjk + htmp vk+1 = vk+1 − htmp vj • hk+1.html. . One can reorthogonalize in every iteration or only if a test [147] detects a loss of orthogonality. .ec-securehost. . hjk = vk+1 vj ii. . 0. one can still lose orthogonality in the columns of V . . vk+1 = vk+1 − hjk vj (c) hk+1. For example.

say.001 0 0   . We employ this test in the MATLAB code gmres with δ = 10−3 . because of its simplicity. by QR factorization or the MATLAB backward division operator. We provide an implementation of Algorithm gmresb in the collection of MATLAB codes. . 1)T .10) Avk 2 + δ vk+1 2 = Avk 2 to working precision. in step 2e of gmresb is O(k3 ) floating-point operations. especially when implemented in an environment like MATLAB. Reorthogonalization is done after the Gram–Schmidt loop and before vk+1 is normalized if (3. For large k. The method based on (3.com/SIAM/FR16. the brute-force method can be very costly. k scalar products. When k is not too large and the cost of matrix-vector products is high. x0 = (0.10) is the default in the MATLAB code gmresa. and the solution of the Hessenberg least squares problem in step 2e.11) . This doubles the cost of the computation of V and is usually unnecessary. however.Copyright ©1995 by the Society for Industrial and Applied Mathematics. While classical Gram–Schmidt fails. A variation on a method from [147] is used in [22]. we find in the MATLAB environment that a solution to full precision requires more than three iterations unless reorthogonalization is applied after every iteration.0011 0  . 0)T . More efficient and equally effective approaches are based on other ideas. Hence the total cost of the m GMRES iterations is m matrix-vector products and O(m4 + m2 N ) floating-point operations. 1. 0. 42 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS One approach to reorthogonalization is to reorthogonalize in every step.ec-securehost.10) is almost as effective as the much more expensive approach of reorthogonalizing in every step. reorthogonalization based on the test (3. To illustrate the effects of loss of orthogonality and those of reorthogonalization we apply GMRES to the diagonal system Ax = b where b = (1.1 we tabulate relative residuals as a function of the iteration counter for classical Gram–Schmidt without reorthogonalization (CGM). This electronic version is for personal use and may not be duplicated or distributed.10) (MGM-PO). and reorthogonalization in every iteration (MGM-FO).html. modified Gram–Schmidt without reorthogonalization (MGM). The k scalar products require O(kN ) floating-point operations and the cost of a solution of the Hessenberg least squares problem. a bruteforce solution to the least squares problem using the MATLAB backward division operator is not terribly inefficient. the reorthogonalization strategy based on (3. Buy this book from SIAM at http://www. The kth GMRES iteration requires a matrix-vector product. and (3. A= 0 0 0 104   While in infinite precision arithmetic only three iterations are needed to solve the system exactly. In Table 3. This is an appealing algorithm. The idea is that if the new vector is very small relative to Avk then information may have been lost and a second pass through the modified Gram–Schmidt process is needed.

ec-securehost. This implementation maintains the QR factorization of Hk in a clever way so that the cost for a single GMRES iteration is O(N k) floating-point operations.42e-08 5. An N × N Givens rotation replaces a 2 × 2 block on the diagonal of the Buy this book from SIAM at http://www. or a shifted Arnoldi process [197] are much more efficient than the brute-force approach in Algorithm gmresb.16e-01 3. π].53e-05 2. The O(k 2 ) cost of the triangular solve and the O(kN ) cost of the construction of xk are incurred after termination. Implementation: Givens rotations If k is large.00e+00 8. The orthogonal matrix G rotates the vector (c.00e-05 2.12) G= .com/SIAM/FR16. GMRES ITERATION Table 3.70e-08 3. implementations using Givens rotations [167].34e-34 3. The implementation in Algorithm gmres and in the MATLAB code collection is from [167].88e-02 6.00e+00 8.42e-08 3.16e-01 3. Householder reflections [195].1 Effects of reorthogonalization.88e-02 6.16e-01 3. G c −s = 1 0 .88e-02 6. This electronic version is for personal use and may not be duplicated or distributed.00e+00 8.74e-05 3. 43 k 0 1 2 3 4 5 6 7 8 9 10 CGM 1.00e+00 8. . which makes an angle of −θ with the x-axis through an angle θ so that it overlaps the x-axis. −s).04e-24 MGM-FO 1.88e-02 6. where c = cos(θ).Copyright ©1995 by the Society for Industrial and Applied Mathematics.html. s = sin(θ) for θ ∈ [−π. [22]. A 2 × 2 Givens rotation is a matrix of the form c −s s c (3.87e-05 3.16e-01 3.04e-18 MGM-PO 1.37e-05 MGM 1.5.69e-05 4.35e-05 3.74e-05 2.

and noting that βe1 − Hk y k Rk+1 = Q(βe1 − Hk y k ) Rk+1 = g − Rk y k Rk+1 . then the first column of H now has only a single nonzero element h11 . .  0 .. G1 we see that QH = R is upper triangular. of course.html. 0 . 44 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS N × N identity matrix with a 2 × 2 Givens rotation. Our notation is that Gj (c. We reduce H to triangular form by first multiplying the matrix by a Givens rotation that annihilates h21 (and. Buy this book from SIAM at http://www. Continuing in this way and setting Q = GN . changes h11 and the subsequent columns).14) c1 = h11 / h2 + h2 and s1 = −h21 / h2 + h2 .   . where Rk is the k + 1 × k triangular factor of the QR factorization of Hk .  .. .. s c        1 0         . 22 32 22 32 and annihilate h32 . This reduction can be accomplished in O(N ) floating-point operations and hence is far more efficient than a solution by a singular value decomposition or a reduction based on Householder transformations. c −s   . .Copyright ©1995 by the Society for Industrial and Applied Mathematics. we can now apply G2 (c2 . .. They are of particular value in reducing Hessenberg matrices to triangular form and thereby solving Hessenberg least squares problems such as the ones that arise in GMRES. s1 ) by (3.ec-securehost.. G= .        (3. s) is an N × N givens rotation of the form (3. We define G1 = G1 (c1 . . Similarly.13) with a 2 × 2 Givens rotation in rows and columns j and j + 1. .  .. A straightforward application of these ideas to Algorithm gmres would solve the least squares problem by computing the product of k Givens rotations Q... [184].com/SIAM/FR16.13) 0 . This electronic version is for personal use and may not be duplicated or distributed. 0 1 0 0 . s2 ) to H. . This method is also used in the QR algorithm for computing eigenvalues [89].. . Givens rotations are used to annihilate single nonzero elements of matrices in reduction to triangular form [89]. setting g = βQe1 . . Let H be an N × M (N ≥ M ) upper Hessenberg matrix with rank M . . 11 21 11 21 If we replace H by G1 H. . Note that G2 does not affect the first column of H.. 1 . where (3.  .15) c2 = h22 / h2 + h2 and s2 = −h32 / h2 + h2 . 0 .

β = ρ. s1 )x. . sk )g. . sk ) . g = ρ(1. (g) ρ = |(g)k+1 |.k = ck hk. If k > 1 apply Qk−1 to the kth column of H. however.k − sk hk+1. . Buy this book from SIAM at http://www. GMRES ITERATION 45 In the context of GMRES iteration. The MATLAB implementation of Algorithm gmres uses (3. .k . We overwrite the upper triangular part of Hk with the triangular part of the QR factorization of Hk in the MATLAB code.1. Set (w)i = (g)i for 1 ≤ i ≤ k. we add the new column hk+2 to Hk . gmres(x.j = hi. . ν = 2 i. ck = hk. we can incrementally perform the QR factorization of H as the GMRES iteration progresses [167].Copyright ©1995 by the Society for Industrial and Applied Mathematics. G1 (c1 . . and finally setting Qk+1 = Gk+1 Qk and forming Rk+1 by augmenting Rk with Gk+1 Qk hk+2 . 0.ec-securehost. ρ) 1.10) to test for loss of orthogonality. 4. r = b − Ax. . . k T i. kmax. .html. h2 + h2 k. then computing the Givens rotation Gk+1 that annihilates the (k + 2)nd element of Qk hk+2 . Algorithm 3. v1 = r/ r 2 . . 0)T ∈ Rkmax+1 2. hk+1. While ρ > b 2 and k < kmax do (a) k = k + 1 (b) vk+1 = Avk for j = 1. ρ = r 2 . xk = x0 + Vk y k .j for 1 ≤ i.com/SIAM/FR16.k k+1.k = vk+1 2 (d) Test for loss of orthogonality and reorthogonalize if necessary.k . This electronic version is for personal use and may not be duplicated or distributed. b.k /ν. after orthogonalization. sk = −hk+1. The MATLAB implementation of Algorithm gmres stores Qk by storing the sequences {cj } and {sj } and then computing the action of Qk on a vector x ∈ Rk+1 by applying Gj (cj . note that if Rk = Qk Hk and. . A. To see this. 3. k = 0. Solve the upper triangular system Ry k = w. vk+1 = vk+1 − hjk vj (c) hk+1. (e) vk+1 = vk+1 / vk+1 (f) ii.k = 0 iv. sj ) in turn to obtain Qk x = Gk (ck . we can update both Qk and Rk by first multiplying hk+2 by Qk (that is applying the first k Givens rotations to hk+2 ).5. j ≤ k. hjk = vk+1 vj ii. g = Gk (ck .k /ν hk. iii. Set ri.

2. m. fail to terminate. it may be impossible to store enough basis vectors and the iteration may have to be restarted.8 we give an example of how this squaring of the condition number can lead to failure of the iteration. A.Copyright ©1995 by the Society for Industrial and Applied Mathematics. therefore. b. kmax. Restarting can seriously degrade performance. Other methods for nonsymmetric systems The methods for nonsymmetric linear systems that receive most attention in this book. Consistently with our implementation of GMRES. k=1 3. A. [74]. . can be analyzed by a common residual polynomial approach. k = m 3. Algorithm 3. and {vk }m . In § 3. The methods we describe in this section fall short of the ideal. Aside from the parameter m. m. be based on some kind of minimization principle or conjugacy. x. This implementation does not test for success and may. CGNR. gmres(x. the storage requirements increase as the iteration progresses. GMRES. and only terminate if an acceptable approximate solution has been found. but. share the properties that they are easy to implement. ρ) 2. A. Hence. and converge in N iterations for all nonsingular A.5. and CGNE. these methods can also be implemented in a manner that incorporates Buy this book from SIAM at http://www. 46 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS We close with an example of an implementation of GMRES(m) . but can still be quite useful.html. However. .ec-securehost. methods based on short-term recurrences such as CG that also satisfy minimization or conjugacy conditions cannot be constructed for general matrices. we take the view that preconditioners will be applied externally to the iteration. An ideal method would. m. as with CG. the arguments to Algorithm gmresm are the same as those for Algorithm gmres.6. ρ) 1. This electronic version is for personal use and may not be duplicated or distributed. ρ) (b) k = k + m The storage costs of m iterations of gmres or of gmresm are the m + 2 vectors b. CGNR and CGNE have the disadvantages that a transpose-vector product must be computed for each iteration and that the coefficient matrix AT A or AAT has condition number the square of that of the original matrix. . b. You are asked to repair this in exercise 7. However.com/SIAM/FR16. We discuss only a small subset of these methods and refer the reader to [12] and [78] for pointers to more of the literature on this subject. For large and ill-conditioned problems. have modest storage requirements that do not depend on the number of iterations needed for convergence. While ρ > b 2 and k < kmax do (a) gmres(x. gmresm(x. . All the methods we present in this section require two matrix-vector products for each iteration. GMRES needs only matrix-vector products and uses A alone. like CG. b. only need matrix-vector products. as we have seen. a basis for Kk must be stored to compute xk .

{ˆk } and search directions {pk }. k = 0 ˆ 2. . Even if these quantities are nonzero but very small. ρ0 = 1. AT r0 . p = r + β p ˆ ˆ ˆ (d) v = Ap (e) α = ρk /(ˆT v) p (f) x = x + αp (g) r = r − αv. Do While r 2 > b 2 and k < kmax (a) k = k + 1 (b) ρk = rT r. If. . the algorithm can become unstable or produce inaccurate results. kmax) ˆ 1. ˆk In the symmetric positive definite case (3.16) where T rk w = 0 for all w ∈ Kk . β = ρk /ρk−1 ˆ (c) p = r + βp.html. Kk = span(ˆ0 . i.6. b. [76]. however.1. If either ρk−1 = 0 or p ˆ place. .ec-securehost. Using the notation of Chapter 2 and [191] we give an implementation of BiCG making the choice r0 = r0 . This electronic version is for personal use and may not be duplicated or distributed. Bi-CG. e. Bi-CG produces the same iterations as CG (but computes everything except x twice). will not ˆ T Ap = 0 we say that a breakdown has taken vanish. [78]. The algorithm gets its name because it produces sequences of residuals {rk }. r = r. A is not spd. r = b − Ax. The earliest such method Bi-CG (Biconjugate gradient) [122].1. This algorithmic description is explained and ˆ motivated in more detail in [122]. and [191]. r0 is a user-supplied vector and is ˆ ˆ often set to r0 . the kth residual must satisfy the bi-orthogonality condition (3. instead. . 3. the denominator in step 2e. . (AT )k−1 r0 ) r ˆ ˆ is the Krylov space for AT and the vector r0 . there is no guarantee that ρk in step 2b or pT Ap. Algorithm 3.6. r = r − αAT p ˆ ˆ ˆ Note that if A = AT is spd. bicg(x. p = p = 0. [76].Copyright ©1995 by the Society for Industrial and Applied Mathematics. does not enforce a minimization principle. A. .2) for CG [89]. GMRES ITERATION 47 the preconditioner and uses the residual for the original system to control termination.16) is equivalent to the minimization principle (2. rk rl = 0 if k = l and the search directions {pk } and {ˆk } satisfy ˆT p the bi-conjugacy property pT Apl = 0 if k = l. We do not recommend use of Bi-CG and present this algorithm only as a basis for discussion of some of the properties of this class of algorithms. While there are approaches Buy this book from SIAM at http://www. {ˆk } such that bi-orthogonality r p holds.com/SIAM/FR16.

17). but not necessarily in terms of computational work). However. 48 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS [80]. The possibility of breakdown is small and certainly should not eliminate the algorithms discussed below from consideration if there is not enough storage for GMRES to perform well. This explains the name. [79]. there are no methods that both limit storage and completely eliminate the possibility of breakdown [74]. 3. One should also keep in mind that a single Bi-CG iteration requires two matrix-vector products and a GMRES iterate only one. Breakdowns aside. will always reduce the residual more rapidly than Bi-CG (in terms of iterations. where pk is the same polynomial produced by Bi-CG and used in ¯ (3. reminding us that GMRES.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Buy this book from SIAM at http://www. ˆ p p r p ˆ The other references to AT can be eliminated in a similar fashion and an iteration satisfying rk = pk (A)2 r0 (3. if sufficient storage is available. Once breakdown has occurred. [81]. A remedy for one of the problems with Bi-CG is the Conjugate Gradient Squared (CGS) algorithm [180]. the effort in computing r ˆ at each iteration is wasted in that r makes no contribution to x. A transposevector product is needed. which at the least will require additional programming and may not be possible at all.ec-securehost. Moreover. [78]. one can restart the Bi-CG iteration or pass the computation to another algorithm. [148].17) rk = pk (A)r0 and rk = pk (AT )ˆ0 . there are other problems with Bi-CG.17) implies that the scalar product rT r in Bi-CG (step 2b of ˆ Algorithm bicg) can be represented without using AT as T rk rk = (¯k (A)r0 )T (¯k (AT )ˆ0 ) = (¯k (A)2 r0 )T r0 .6. Conjugate Gradient Squared.18) ¯ is produced. We can compare the best-case performance of Bi-CG with that of GMRES. but that the cost of the GMRES iteration increases (in terms of floating-point operations) as the iteration progresses. such as GMRES. This electronic version is for personal use and may not be duplicated or distributed. The algorithm takes advantage of the fact that (3. Note that there is pk ∈ Pk such that both ¯ (3.html.com/SIAM/FR16. by the minimization property for GMRES GM rk RES 2 Bi−CG ≤ rk 2. The performance of the algorithm can be erratic or even unstable with residuals increasing by several orders of magnitude from one iteration to the next. All of the methods we present in this section can break down and should be implemented with that possibility in mind. CGS. ¯ ˆ ¯ r Hence. . to avoid some breakdowns in many variations of Bi-CG. ˆ Bi-CG sometimes performs extremely well and the remaining algorithms in this section represent attempts to capture this good performance and damp the erratic behavior when it occurs.2.

[77]. If the algorithm does not break down.6. Since p2 ∈ P2k we have ¯k (3. CGS ˆ replaces the transpose-vector product with an additional matrix-vector product and applies the square of the Bi-CG polynomial to r0 to produce rk . This may. v0 = Ap0 . r0 = r0 ˆ 3. [191]. GMRES ITERATION 49 The work used in Bi-CG to compute r is now used to update x. βk = ρk /ρk−1 ˆT uk = rk + βk qk pk = uk + βk (qk + βk pk−1 ) vk = Apk Breakdowns take place when either ρk−1 or σk−1 vanish.2.17) and (3. of course. Bi-CGSTAB is the most effective representative of this class of algorithms. ρ0 = r0 r0 . b.21) ¯ ¯ qk (z) = pk (z) + βk qk−1 (z). ¯ We provide no MATLAB implementation of Bi-CG or CGS. that if pk is the residual polynomial from (3. cgs(x.4.6. One can show [180]. change the convergence properties for the worse and either improve good convergence or magnify erratic behavior [134]. Algorithm 3. We present an algorithmic description of CGS that we will refer to in our discussion of TFQMR in § 3.com/SIAM/FR16. CGS has the same potential for breakdown as Bi-CG. A. This description is taken from [77]. . p0 = u0 = r0 = b − Ax 2. This electronic version is for personal use and may not be duplicated or distributed.19) GM r2k RES 2 CGS ≤ rk 2. .html. Buy this book from SIAM at http://www. then αk−1 = 0 for all k.18) then ¯ (3. kmax) 1.Copyright ©1995 by the Society for Industrial and Applied Mathematics.20) ¯ ¯ pk (z) = pk−1 (z) − αk−1 z qk−1 (z) ¯ ¯ where the auxiliary polynomials qk ∈ Pk are given by q0 = 1 and for k ≥ 1 by ¯ (3. In our description of CGS we will index the vectors and scalars that would be overwritten in an implementation so that we can describe some of the ideas in TFQMR later on.ec-securehost. x0 = x. Do While r 2 > b 2 and k < kmax (a) k = k + 1 ˆT (b) σk−1 = r0 vk−1 (c) αk−1 = ρk−1 /σk−1 (d) qk = uk−1 − αk−1 vk−1 (e) xk = xk−1 + αk−1 (uk−1 + qk ) rk = rk−1 − αk−1 A(uk−1 + qk ) (f) ρk = r0 rk . k = 0 ˆT 4.

(3. b.6. Bi-CGSTAB. r0 = r = r. kmax) ˆ ˆT 1. r = b − Ax.8. The reason for this is that the cost of orthogonalization in GMRES can be much more than that of a matrix-vector product if the dimension of the Krylov space is large. 50 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 3. ρk+1 = −ωˆ0 t 2 (h) x = x + αp + ωs (i) r = s − ωt Note that the iteration can break down in steps 2b and 2e. The performance in cases where the spectrum has a significant imaginary part can be improved by constructing qk to have complex conjugate pairs of roots. r. letting s overwrite r when needed. k = 0.3.6.3. The Bi-conjugate gradient stabilized method (BiCGSTAB) [191] attempts to smooth the convergence of CGS by replacing (3. Buy this book from SIAM at http://www. which can be thought of [12] as a blend of Bi-CG and GMRES(1).19). ρ0 = α = ω = 1. The examples in [191] indicate the effectiveness of this approach. t). A single iteration reˆ quires four scalar products. The cost in storage and in floating-point operations per iteration remains bounded for the entire iteration. There is no convergence theory for Bi-CGSTAB and no estimate of the residuals that is better than that for CGS.18) with rk = qk (A)pk (A)r0 (3. p. which is described and motivated fully in [191].22) where qk (z) = k i=1 (1 − ωi z). A. This electronic version is for personal use and may not be duplicated or distributed. Bi-CGSTAB can have a much lower average cost per iterate than GMRES.com/SIAM/FR16. we use r0 = r0 ˆ and follow the notation of [191]. We provide an implementation of Algorithm bicgstab in the collection of MATLAB codes. We will present such a case in § 3. . v.html. bicgstab(x. One must store seven vectors (x. v = p = 0. In a situation where many GMRES iterations are needed and matrix-vector product is fast. ρ1 = r0 r ˆ 2.ec-securehost. The constant ωi is selected to minimize ri = qi (A)pi (A)r0 as a function of ωi . In our description of the implementation. t = As rT (g) ω = tT s/ t 2 . [98]. Do While r 2 > b 2 and k < kmax (a) k = k + 1 (b) β = (ρk /ρk−1 )(α/ω) (c) p = r + β(p − ωv) (d) v = Ap rT (e) α = ρk /(ˆ0 v) (f) s = r − αv. . r0 . Algorithm 3. b.Copyright ©1995 by the Society for Industrial and Applied Mathematics.

 0 . [202]. We express (3. [77]. if m = 2k is even. We will focus on the transposefree algorithm TFQMR proposed in [77]. therefore. All algorithms in this family minimize the norm of an easily computable quantity q = f − Hz.20) and (3. ..4. α0 . the p CGS residual satisfies CGS rk = w2k+1 . The quasi-residual is related to the true residual by a full-rank linear transformation r = Lq and reduction of the residual can be approximately measured by reduction in q. and Bm is the (m + 1) × m matrix given by j=1    −1   Bm =  0   . −1 . .. Now we consider the quasi-minimal residual (QMR) family of algorithms [80]. .24) where Ym is the N × m matrix with columns {yj }m . . This electronic version is for personal use and may not be duplicated or distributed. Buy this book from SIAM at http://www.  .com/SIAM/FR16..   1  Our assumptions that no breakdown takes place imply that Km is the span of {yj }m and hence j=1 xm = x0 + Ym z (3. If we let r be the nearest integer less than or equal to a real r we have Aym = (wm − wm+1 )/α (m−1)/2 . . α (m−1)/2 ) . if m = 2k is even. 1 0 1 . . a quasi-residual over z ∈ RN .   (¯k (A))2 r0  p   p (A)¯ ¯k pk−1 (A)r0 q Here the sequences {¯k } and {¯k } are given by (3.26) rm = r0 − AYm z = Wm+1 (e1 − Bm z). The specific formulation of q and L depends on the algorithm. 0 . .23) in matrix form as AYm = Wm+1 Bm . . [37]. [81]. . to illustrate the quasi-minimization idea.  .6. (3. . (3. if m = 2k − 1 is odd. Returning to Algorithm cgs we define sequences ym = and wm =   uk−1    q k if m = 2k − 1 is odd.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Therefore.  .ec-securehost. . .23) where the denominator is nonzero by assumption.html. αk = 0 for all k ≥ 0. GMRES ITERATION 51 3.25) for some z ∈ Rm . Wm+1 the N × (m + 1) j=1 matrix with columns {wj }m+1 . TFQMR. ..21). Hence. . We assume that CGS does not break down and. (3.  . 0 −1   −1  0  diag(α0 . . .

also use the quasi-minimization condition to estimate the TFQMR residual in terms of residual polynomials. As one can see from that description. Let A be an N × N matrix and let x0 . 0)T ∈ Rm . Hence (3. as in our discussion of GMRES. Theorem 3. . We can.31) rm 2 2 then the columns of Wm+1 Ω−1 have m+1 2 = Wm+1 Ω−1 (fm+1 − Hm z) m+1 √ ≤ τm m + 1.1. therefore. Instead we have the quasi-residual norm τm = fm+1 − Hm zm 2 . The quasi-minimization idea is to solve (3. The solution to the least squares problem (3. Assume that the TFQMR iteration does not break down or terminate with GM the exact solution for k iterations and let 1 ≤ m ≤ 2k. [78]. Let ξm be the smallest singular value of Wm+1 Ω−1 . If Wm+1 Ω−1 were an orthogonal matrix. Then if ξm > 0 m+1 (3. the conditioning of the matrix Wm+1 can be improved by scaling.Copyright ©1995 by the Society for Industrial and Applied Mathematics.28) rm = r0 − AYm z = Wm+1 Ω−1 (fm+1 − Hm z) m+1 Ω = diag(ω1 . . Let (3. . .29) can be done by using Givens rotations to update the factorization of the upper Hessenberg matrix Hm [77] and this is reflected in the implementation description below. each k produces two approximate solutions (corresponding to m = 2k − 1 and m = 2k). This electronic version is for personal use and may not be duplicated or distributed. Buy this book from SIAM at http://www.29) minimizez∈Rm fm+1 − Hm z 2 . 0. ω2 . we could minimize rm by solving m+1 the least squares problem (3.ec-securehost. One can [80]. . The diagonal entries in the matrix Ωm+1 are called weights. Now if we pick the weights ωi = wi norm one. base our termination criterion on τm .com/SIAM/FR16. . and we do this in Algorithm tfqmr.32) GM τm ≤ rm RES 2 /ξm . 52 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS where.26) as (3. approximate residuals are not computed in Algorithm tfqmr. In a sense. . ωm ) where fm+1 = Ωm+1 e1 = ωm+1 e1 and Hm = Ωm+1 Bm is bidiagonal and hence upper Hessenberg. The weights make it possible to derive straightforward estimates of the residual in terms of easily computed terms in the iteration.29) (thereby quasi-minimizing the residual) to obtain zm and then set (3.6. . Note that if k corresponds to the iteration counter for CGS.30) xm = x0 + Ym zm . .html. Let rm RES be the residual for the mth GMRES iteration.27) and rewrite (3. e1 = (1. b ∈ RN be given.

Let A be an N × N matrix and let x0 . u1 = v = Ay1 . w1 = y1 = r0 = b − Ax. Then. We may write GM GM rm RES = r0 + Ym zm RES and obtain GM rm RES 2 GM = Wm+1 Ω−1 (fm+1 −Hm zm RES ) m+1 2 GM ≥ ξm fm+1 −Hm zm RES 2. The implementation below follows [77] with r0 = r0 used throughout. For 1 ≤ m ≤ 2k let ξm be the smallest singular value of Wm+1 Ω−1 m+1 and let xm be given by (3. θ = 0. η = 0 2. 2 (m = 2k − 2 + j) w = w − αk−1 uj d = yj + (θ2 η/αk−1 )d √ θ = w 2 /τ . A. .2. iii. ˆ Algorithm 3.4.32) and use (3. This electronic version is for personal use and may not be duplicated or distributed. j = 1. Do While k < kmax (a) k = k + 1 T (b) σk−1 = r0 v. Then within (N + 1)/2 iterations the TFQMR iteration will either break down or terminate with the solution. GMRES ITERATION 53 Proof. As a corollary we have a finite termination result and a convergence estimate for diagonalizable matrices similar to Theorem 3. d = 0 T ρ0 = r0 r0 .1.1. if ξm > 0. c = 1/ 1 + θ2 τ = τ θc. y2 = y1 − αk−1 v.3 to (3.30). α = ρk−1 /σk−1 . vi. We apply Theorem 3.1. iv.6. b ∈ RN be given. ii. η = c2 αk−1 x = x + ηd √ If τ m + 1 ≤ b 2 terminate successfully T (d) ρk = r0 w.com/SIAM/FR16. β = ρk /ρk−1 Buy this book from SIAM at http://www. b.ec-securehost. Corollary 3. Assume that TFQMR does not break down or terminate with the solution for k iterations. tfqmr(x. τ = r 2 . Theorem 3. k = 0. Hence. kmax) 1. u2 = Ay2 (c) For i.31) to obtain the following result.6. by the quasi-minimization property GM rm RES 2 ≥ ξm τm as desired.6. rm r0 2 2 ≤ √ −1 m + 1ξm κ(V ) max |φ(z)| z∈σ(A) for all φ ∈ Pm .3. .html.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Let A = V ΛV −1 be a nonsingular diagonalizable matrix. v.

We take advantage of this in the MATLAB code tfqmr to potentially reduce the matrix-vector product cost by one. [34]. the right-hand side vector b. As in § 2. As before. For the computations reported in this section we took a1 (x. y) = f (x. CGNR. y) on 0 < x. y < 1. a MATLAB function for the matrix-vector product. u1 = Ay1 (f) v = u1 + β(u2 + βv) Note that y2 and u2 = Ay2 need only be computed if the loop in step 2c does not terminate when j = 1. [121]. The motivation for this choice is similar to that for the conjugate gradient computations and has been partially explained theoretically in [33]. y < 1 subject to homogeneous Dirichlet boundary conditions u(x. y) = −(uxx (x. The inputs. . xj ) where xi = ih for 1 ≤ i ≤ n. and iteration parameters to specify the maximum number of iterations and the termination criterion. 0) = u(x. y) = 1.8 × 10−4 b 2 . y) +a2 (x. BiCGSTAB. whose storage requirements do not increase with the number of iterations. We let n = 31 to create a system with 961 unknowns.7. 54 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS (e) y1 = w + βy2 . The unknowns are uij ≈ u(xi . 0 < x. and a3 (x. For methods like GMRES(m). If we let Gu denote the action of the Poisson solver on u.Copyright ©1995 by the Society for Industrial and Applied Mathematics. rather we pass the preconditioned problem.33) (Lu)(x. a2 (x.com/SIAM/FR16. y) = 1. 3. As a preconditioner we use the fast Poisson solver fish2d described in § 2. This electronic version is for personal use and may not be duplicated or distributed. y) + uyy (x. We consider the discretization of the partial differential equation (3. the preconditioned system is GLu = Gf . we allow for more iterations. y) = u(1. We do not pass the preconditioner to the GMRES code. y)u(x. For general coefficients {aj } the operator L is not self-adjoint and its discretization is not symmetric. Buy this book from SIAM at http://www. We compute Lu in a matrix-free manner as we did in § 2. described in the comment lines. y) + a3 (x.html. Examples for GMRES iteration These examples use the code gmres from the collection of MATLAB codes.ec-securehost. y) = 20y. y)uy (x. 1) = u(0.7. and TFQMR. are the initial iterate x0 . we limit the number of GMRES iterations to 60. y)ux (x. y) = 0. we expect second-order accuracy and terminate the iteration when rk 2 ≤ h2 b 2 ≈ 9.7 we discretize with a five-point centered difference scheme with n2 points and mesh width h = 1/(n + 1). In all the GMRES examples. y)) + a1 (x.7. and [139].

For elliptic problems like (3. However. In other computing environments preconditioners which vectorize or parallelize particularly well might perform in the same way. To apply CGNR to the preconditioned problem. The cost of the application of the transpose of L or GL is the same as that of the application of L or GL itself.1 and 3. In Figs. In both figures the solid line is the history of the preconditioned iteration and the dashed line that of the unpreconditioned. we require the transpose of GL. Hence.7 million floating point operations and the preconditioned iteration required 13 iterations and 2.6 million floating point operations. CGNR has the effect of squaring the condition number. The reason for this is that the FFT routine is a MATLAB intrinsic and is not interpreted MATLAB code. in the case where the cost of the matrix-vector product dominates the computation.html.3 shows. This electronic version is for personal use and may not be duplicated or distributed. Note that the plots are semilog plots of r 2 / b 2 and that the right-hand sides in the preconditioned and unpreconditioned cases are not the same. a single CGNR iteration costs roughly the same as two GMRES iterations. which is given by (GL)∗ u = L∗ G∗ u = L∗ Gu. even in the unpreconditioned case. the restarted iteration terminated successfully. we restarted the algorithm after three iterations.Copyright ©1995 by the Society for Industrial and Applied Mathematics. the transpose (adjoint) of L is given by L∗ u = −uxx − uyy − a1 ux − a2 uy + a3 u as one can derive by integration by parts. and TFQMR iteration When a transpose is available. The execution times in our environment indicate that the preconditioned iterations are even faster than the difference in operation counts indicates.4. Examples for CGNR.com/SIAM/FR16. In this case the MATLAB flops command indicates that the unpreconditioned iteration terminated after 223 iterations and 9. Note the importance of preconditioning. 3. GMRES(3) in the terminology of § 3. For the restarted iteration. . The preconditioned iteration required 1.ec-securehost. Bi-CGSTAB. 3. However.8. 3.5 ).2 we plot iteration histories corresponding to preconditioned or unpreconditioned GMRES. CGNR (or CGNE) is an alternative to GMRES that does not require storage of the iteration history. The dashed line corresponds to Buy this book from SIAM at http://www. GMRES ITERATION 55 u0 = 0 was the initial iterate. As in § 2. The MATLAB flops command indicates that the unpreconditioned iteration required 7. this effect has serious consequences as Fig.3 million floating point operations and terminated after 8 iterations.33).7 we took the right hand side so that the exact solution was the discretization of 10xy(1 − x)(1 − y) exp(x4. Restarting after 3 iterations is more costly in terms of the iteration count than not restarting.6 million floating point operations and converged in 56 iterations.

This is a result of squaring the large condition number.e.1.ec-securehost. The solid line corresponds to the CGNR applied to the preconditioned problem. CGNR on the unpreconditioned problem.com/SIAM/FR16.. i..e. 56 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 0 10 10 Relative Residual -1 10 -2 10 -3 10 -4 0 10 20 30 Iterations 40 50 60 Fig. 3. We limited the number of iterations to 310 = 10m and the unpreconditioned formulation of CGNR had made very little progress after 310 iterations. GMRES(3) for 2-D elliptic equation. i. CG applied to L∗ G2 Lu = L∗ G2 f . This electronic version is for personal use and may not be duplicated or distributed. converging in eight iterations.2. The matrix corresponding to the discretization of L∗ L has a large condition number.7 million floating-point operations Buy this book from SIAM at http://www. The preconditioned formulation did quite well.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 3. 10 0 10 Relative Residual -1 10 -2 10 -3 10 -4 0 50 100 Iterations 150 200 250 Fig.html. GMRES for 2-D elliptic equation. . The MATLAB flops command indicates that the unpreconditioned iteration required 13. CG applied to L∗ Lu = L∗ f .

possibly. GMRES ITERATION 57 and the preconditioned iteration 2.9. see Exercise 3.4. with the unpreconditioned iteration being very irregular.3. for the final iterate.ec-securehost. CGNR for 2-D elliptic equation.5 we plot the convergence histories in terms of the approximate residual 2k + 1τ2k given by (3. We applied Bi-CGSTAB to both the preconditioned and unpreconditioned forms of (3.9. the behavior of even the preconditioned iteration can also be poor if the coefficients of the first derivative terms are too large.33) with the same data and termination criterion as for the GMRES example. 3. Neither convergence history is monotonic. . the solid line corresponds to the preconditioned iteration and the dashed line to the unpreconditioned iteration. 10 0 10 Relative Residual -1 10 -2 10 -3 10 -4 0 50 100 150 200 Iterations 250 300 350 Fig. and the graphs are monotonic. The preconditioned iteration terminated in six iterations (12 matrix-vector products) and required roughly 1. In Fig. However. This electronic version is for personal use and may not be duplicated or distributed. since our unpreconditioned matrix-vector product is so inexpensive. This code has the same input/output arguments as gmres.html. We plot the iteration histories in Fig. As in the previous examples. We use the code tfqmr that is included in our collection. We repeat the experiment with TFQMR as our linear solver.√3.7 million floating-point operations.9.8.4 million. CGNE has similar performance.31).2 million floating-point operations.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 3. The examples for Bi-CGSTAB use the code bicgstab from the collection of MATLAB codes. The approximate residual is given in terms of the quasi-residual norm τm rather than the true residual. The unpreconditioned iteration took 40 iterations and 2.com/SIAM/FR16. a better cost/iterate than GMRES in the unpreconditioned case. We see a considerable improvement in cost over CGNR and. Our Buy this book from SIAM at http://www. but not varying by many orders of magnitude as a CGS or Bi-CG iteration might. as you are asked to investigate in Exercise 3. We plot only full iterations (m = 2k) except.

The unpreconditioned iteration required roughly 4.com/SIAM/FR16.Copyright ©1995 by the Society for Industrial and Applied Mathematics.9 million floating-point operations and 7 iterations.1 million floating-point operations and 67 iterations for termination. 3.31) to make termination decisions. termination criterion is based on the quasi-residual and (3. The preconditioned iteration was much more efficient. This led to actual relative residuals of 7 × 10−5 for the unpreconditioned iteration and 2 × 10−4 for the preconditioned iteration. stopping the iteration when τm m + 1/ b 2 ≤ h−2 .6. This electronic version is for personal use and may not be duplicated or distributed.31) as described in √ § 3. 58 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 1 10 10 0 Relative Residual 10 -1 10 -2 10 -3 10 -4 0 5 10 15 20 Iterations 25 30 35 40 Fig.4. Bi-CGSTAB for 2-D elliptic equation. The plateaus in the graph of the convergence history for the Buy this book from SIAM at http://www. indicating that it is reasonable to use the quasi-residual estimate (3. taking 1. 10 0 10 Relative Approximate Residual -1 10 -2 10 -3 10 -4 10 -5 0 10 20 30 40 Iterations 50 60 70 Fig.html. .5.4.ec-securehost. TFQMR for 2-D elliptic equation. 3.

.html. Buy this book from SIAM at http://www.ec-securehost.Copyright ©1995 by the Society for Industrial and Applied Mathematics. GMRES ITERATION 59 unpreconditioned iteration are related (see [196]) to spikes in the corresponding graph for a CGS iteration. This electronic version is for personal use and may not be duplicated or distributed.com/SIAM/FR16.

7.9. CGNE. 0  0 . . Let A = Jλ be an elementary Jordan block of size N corresponding to an eigenvalue λ = 0.3. . Bi-CGSTAB.. Give an example of a matrix that is not diagonalizable.9.4 and Theorem 3. 3.9. (M f )(0) = (M f )(1) = 0. and λ ∈ R such that the GMRES residuals satisfy rk 2 = r0 2 for all k < N . That is.1 and add two more columns corresponding to classical Gram–Schmidt orthogonalization with reorthogonalization at every step and determine when the test in (3.9.1.ec-securehost.html. How does the performance of the iteration depend on the mesh? Try other preconditioners such as Gauss–Seidel and Jacobi. Consider the centered difference discretization of −u + u + u = 1. Try this for meshes with 50. b ∈ RN .4. This means that    Jλ =      λ  0  1 λ .8).9. ... How do other methods such as CGNR. Buy this book from SIAM at http://www. Prove that the Arnoldi process (unless it terminates prematurely with a solution) produces a basis for Kk+1 that satisfies (3. Modify the MATLAB GMRES code to allow for restarts. Exercises on GMRES 3. points. 3. 3.1. 60 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 3. .Copyright ©1995 by the Society for Industrial and Applied Mathematics. Solve this problem with GMRES (without preconditioning) and then apply as a preconditioner the map M defined by −(M f ) = f. One way to do this is to call the gmres code with another code which has an outer loop to control the restarting as in Algorithm gmresm.9.2. . 100.9. How would you test for failure in a restarted algorithm? What kinds of failures can occur? Repeat Exercise 6 with a limit of three GMRES iterations before a restart..6.com/SIAM/FR16. . precondition with a solver for the high-order term in the differential equation using the correct boundary conditions. and TFQMR perform? 3.1. 1 λ 0 λ     . 0 1 . Duplicate Table 3. Prove Theorem 3. .  0   1  Give an example of x0 . . This electronic version is for personal use and may not be duplicated or distributed. λ 0 . 3.5.5. u(0) = u(1) = 0. 200. See [167] for another example.9. How do your conclusions compare with those in [160]? 3.10) detects loss of orthogonality.

Are the solutions to (3.7 and § 3.7 and § 3.8 equally accurate? Compare the final results with the known solution. This electronic version is for personal use and may not be duplicated or distributed. Buy this book from SIAM at http://www.8. .7.9. 3.8.10.9. Experiment with different values of m.9. Why is the performance of the methods sensitive to ay and ax ? 3. Consider preconditioning from the right. Why might GMRES(m) be faster than GMRES? Vary the dimension and the coefficient ay . GMRES ITERATION 61 3. Implement Bi-CG and CGS and try them on some of the examples in this chapter.9.12. Duplicate the results in § 3. For the PDE in § 3.ec-securehost. Try ay = 5y. apply GMRES.7.html.33) in § 3. Bi-CGSTAB.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Compare timings and solution accuracy with those obtained by left preconditioning. and/or CGNE. Compare the execution times in your computing environment. How does the performance differ from CGNR and GMRES? 3. CGNR.11. to the equation LGw = f and then recover the solution from u = Gw. TFQMR.9. 3.com/SIAM/FR16. Apply CGNE to the PDE in § 3.9.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

62

ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

Chapter 4

Basic Concepts and Fixed-Point Iteration

This part of the book is about numerical methods for the solution of systems of nonlinear equations. Throughout this part, we will let · denote a norm on RN as well as the induced matrix norm. We begin by setting the notation. We seek to solve (4.1) F (x) = 0. Here F : RN → RN . We denote the ith component of F by fi . If the components of F are differentiable at x ∈ RN we define the Jacobian matrix F (x) by ∂fi F (x)ij = (x). ∂xj The Jacobian matrix is the vector analog of the derivative. We may express the fundamental theorem of calculus as follows. Theorem 4.0.1. Let F be differentiable in an open set Ω ⊂ RN and let ∗ ∈ Ω. Then for all x ∈ Ω sufficiently near x∗ x F (x) − F (x∗ ) = 4.1. Types of convergence Iterative methods can be classified by the rate of convergence. Definition 4.1.1. Let {xn } ⊂ RN and x∗ ∈ RN . Then • xn → x∗ q-quadratically if xn → x∗ and there is K > 0 such that xn+1 − x∗ ≤ K xn − x∗ 2 . • xn → x∗ q-superlinearly with q-order α > 1 if xn → x∗ and there is K > 0 such that xn+1 − x∗ ≤ K xn − x∗ α . • xn → x∗ q-superlinearly if
n→∞ 1 0

F (x∗ + t(x − x∗ ))(x − x∗ ) dt.

lim

xn+1 − x∗ = 0. xn − x∗
65

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

66

ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS

• xn → x∗ q-linearly with q-factor σ ∈ (0, 1) if xn+1 − x∗ ≤ σ xn − x∗ for n sufficiently large. Definition 4.1.2. An iterative method for computing x∗ is said to be locally (q-quadratically, q-superlinearly, q-linearly, etc.) convergent if the iterates converge to x∗ (q-quadratically, q-superlinearly, q-linearly, etc.) given that the initial data for the iteration is sufficiently good. Note that a q-superlinearly convergent sequence is also q-linearly convergent with q-factor σ for any σ > 0. A q-quadratically convergent sequence is q-superlinearly convergent with q-order 2. In general, a q-superlinearly convergent method is preferable to a q-linearly convergent one if the cost of a single iterate is the same for both methods. We will see that often a method that is more slowly convergent, in terms of the convergence types defined above, can have a cost/iterate so low that the slow iteration is more efficient. Sometimes errors are introduced into the iteration that are independent of the errors in the approximate solution. An example of this would be a problem in which the nonlinear function F is the output of another algorithm that provides error control, such as adjustment of the mesh size in an approximation of a differential equation. One might only compute F to low accuracy in the initial phases of the iteration to compute a solution and tighten the accuracy as the iteration progresses. The r-type convergence classification enables us to describe that idea. Definition 4.1.3. Let {xn } ⊂ RN and x∗ ∈ RN . Then {xn } converges to x∗ r-(quadratically, superlinearly, linearly) if there is a sequence {ξn } ⊂ R converging q-(quadratically, superlinearly, linearly) to zero such that xn − x∗ ≤ ξn . We say that {xn } converges r-superlinearly with r-order α > 1 if ξn → 0 q-superlinearly with q-order α. 4.2. Fixed-point iteration Many nonlinear equations are naturally formulated as fixed-point problems (4.2) x = K(x) where K, the fixed-point map, may be nonlinear. A solution x∗ of (4.2) is called a fixed point of the map K. Such problems are nonlinear analogs of the linear fixed-point problems considered in Chapter 1. In this section we analyze convergence of fixed-point iteration, which is given by (4.3) xn+1 = K(xn ).

This iterative method is also called nonlinear Richardson iteration, Picard iteration, or the method of successive substitution.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

2. k ≥ 0.ec-securehost.1. .1. Let Ω ⊂ RN . . Definition 4.2. Compare the proof to that of Lemma 1.2. Proof.Copyright ©1995 by the Society for Industrial and Applied Mathematics. y ∈ Ω. Theorem 4.2.2. xn+k − xn = K(xn+k−1 ) − K(xn−1 ) ≤ γ xn+k−1 − xn−1 ≤ γ K(xn+k−2 ) − K(xn−2 ) ≤ γ 2 xn+k−2 − xn−2 ≤ . Let Ω be a closed subset of RN and let K be a contraction mapping on Ω with Lipschitz constant γ < 1 such that K(x) ∈ Ω for all x ∈ Ω. G is Lipschitz continuous on Ω with Lipschitz constant γ if G(x) − G(y) ≤ γ x − y for all x. Let x0 ∈ Ω. This electronic version is for personal use and may not be duplicated or distributed. ≤ γ n xk − x0 ≤ γ n x1 − x0 /(1 − γ).html. If K has two fixed points x∗ and y ∗ in Ω. . Let Ω ⊂ RN and let G : Ω → RM . .k→∞ and therefore the sequence {xn } is a Cauchy sequence and has a limit x∗ [162].3) converges q-linearly to x∗ with q-factor γ for all initial iterates x0 ∈ Ω.2. x∗ ∈ Ω. Hence lim xn+k − xn = 0 n. and the iteration defined by (4. then x∗ − y ∗ = K(x∗ ) − K(y ∗ ) ≤ γ x∗ − y ∗ . Definition 4. for all n. ≤ γ i x1 − x0 .1. K : Ω → RN is a contraction mapping on Ω if K is Lipschitz continuous on Ω with Lipschitz constant γ < 1. . and therefore xn − x0 = ≤ n−1 i=0 xi+1 n−1 i=0 − xi n−1 i i=0 γ xi+1 − xi ≤ x1 − x0 ≤ x1 − x0 /(1 − γ). The sequence {xn } remains bounded since for all i ≥ 1 xi+1 − xi = K(xi ) − K(xi−1 ) ≤ γ xi − xi−1 . Note that {xn } ⊂ Ω because x0 ∈ Ω and K(x) ∈ Ω whenever x ∈ Ω. Now.1 and Corollary 1. Then there is a unique fixed point of K. FIXED-POINT ITERATION 67 Before discussing convergence of fixed-point iteration we make two definitions.com/SIAM/FR16. The standard result for fixed-point iteration is the Contraction Mapping Theorem [9]. Buy this book from SIAM at http://www.

. 1. y K(x) − K(y) = h|f (x) − f (y)| ≤ hM |x − y|. y(t0 ) = y0 and its solution by the backward Euler method with time step h. i. The standard assumptions We will make the standard assumptions on F . Nonlinear variations of the classical stationary iterative methods such as Gauss–Seidel have also been used.3. One simple application of this theorem is to numerical integration of ordinary differential equations by implicit methods.ec-securehost. and the methods based on Newton’s method which we present in the subsequent chapters are used much more often [83]. see [145] for a more complete discussion of these algorithms. 4. This electronic version is for personal use and may not be duplicated or distributed.3.4) y k denotes the approximation of y(t0 + kh). which completes the proof. e.1.com/SIAM/FR16.4) y n+1 − y n = hf (y n+1 ). say |f (y)| ≤ M and |f (x) − f (y)| ≤ M |x − y| for all x. For simplicity we assume that f is bounded and Lipschitz continuous.html. Assumption 4. For example. Buy this book from SIAM at http://www. This time step may well be too small to be practical. Hence the fixed point is unique. F : Ω → RN ×N is Lipschitz continuous with Lipschitz constant γ. 68 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Since γ < 1 this can only hold if x∗ − y ∗ = 0. (4. 2. y. If hM < 1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. This type of analysis was used by Picard to prove existence of solutions of initial value problems [152]. Finally we note that xn+1 − x∗ = K(xn ) − K(x∗ ) ≤ γ xn − x∗ .1 has a solution x∗ . Advancing in time from y n to y n+1 requires solution of the fixed-point problem y = K(y) = y n + hf (y). In (4. we may apply the contraction mapping theorem with Ω = R since for all x. consider the initial value problem y = f (y). x∗ = y ∗ . From this we conclude that for h sufficiently small we can integrate forward in time and that the time step h can be chosen independently of y. Equation 4.

If xn is the nth iterate of a sequence. F (x)−1 ≤ 2 F (x∗ )−1 .com/SIAM/FR16. This will follow from δ< since then (4.1 to obtain F (x) ≤ 1 0 F (x∗ + te) e dt ≤ 2 F (x∗ ) e Buy this book from SIAM at http://www. FIXED-POINT ITERATION 69 3. Hence (4.6) and (4.5) holds if γδ < F (x∗ ) . Exercise 5.1. we will always denote a root of F by x∗ . . The notation introduced above will be used consistently. Let δ be small enough so that B(δ) ⊂ Ω.5) and Theorem 4.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Lemma 4. where e = x − x∗ . en = xn − x∗ is the error in that iterate. We use (4. The next result (4. we note that if x ∈ B(δ) then x∗ + te ∈ B(δ) for all 0 ≤ t ≤ 1.1 illustrates one way to weaken the standard assumptions. These assumptions can be weakened [108] without sacrificing convergence of the methods we consider here.ec-securehost.0.1 is an important consequence of the standard assumptions.8) I − F (x∗ )−1 F (x) = F (x∗ )−1 (F (x∗ ) − F (x)) ≤ γ F (x∗ )−1 e ≤ γδ F (x∗ )−1 < 1/2. −1 e /2 ≤ F (x) ≤ 2 F (x∗ ) e . We let B(r) denote the ball of radius r about x∗ B(r) = {x | e < r}.html.7) F (x∗ )−1 F (x) ≤ 2 F (x∗ ) . For all x ∈ B(δ) we have F (x) ≤ F (x∗ ) + γ e . Lemma 4. F (x∗ ) is nonsingular. Throughout this part.3.5) (4. Assume that the standard assumptions hold.7. Proof. F (x∗ )−1 2γ −1 To prove the final inequality (4. Then there is δ > 0 so that for all x ∈ B(δ) (4. However the classical result on quadratic convergence of Newton’s method requires them and we make them throughout. This electronic version is for personal use and may not be duplicated or distributed.7).6) is a direct consequence of the Banach Lemma if I − F (x∗ )−1 F (x) < 1/2.3.

which completes the proof. .Copyright ©1995 by the Society for Industrial and Applied Mathematics.com/SIAM/FR16.7). by (4.html. e /2 ≤ F (x∗ )−1 F (x) ≤ F (x∗ )−1 F (x) .ec-securehost. This electronic version is for personal use and may not be duplicated or distributed. 70 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS which is the right inequality in (4. I − F (x∗ )−1 F (x∗ + te)dt ) ≥ e /2.7) note that F (x∗ )−1 F (x) = F (x∗ )−1 =e− and hence.8) F (x∗ )−1 F (x) ≥ e (1 − Therefore 1 0 1 0 1 0 F (x∗ + te)e dt (I − F (x∗ )−1 F (x∗ + te))e dt. Buy this book from SIAM at http://www. To prove the left inequality in (4.

Newton’s method is x+ = xc − F (xc )−1 F (xc ). Chapter 5 Newton’s Method 5.1. This electronic version is for personal use and may not be duplicated or distributed.html. [156].com/SIAM/FR16. The convergence result on Newton’s method follows from Lemma 4.2) Proof. Theorem 5. Let the standard assumptions hold.1. .3.1 hold.2) with K = γ F (x∗ )−1 . By Lemma 4. [13].1) satisfies e+ ≤ K ec 2 . 71 Buy this book from SIAM at http://www. In this language.Copyright ©1995 by the Society for Industrial and Applied Mathematics.ec-securehost.1 and the Lipschitz continuity of F e+ ≤ (2 F (x∗ )−1 )γ ec 2 /2.1. Let the standard assumptions hold. By Theorem 4.3. (5. In the context of single systems this method appeared in the 17th century [140].3. (5. Let δ be small enough so that the conclusions of Lemma 4. Local convergence of Newton’s method We will describe iterative methods for nonlinear equations in terms of the transition from a current iterate xc to a new iterate x+ . The proof of convergence of the complete Newton iteration will be complete if we reduce δ if needed so that Kδ < 1.1.2. [137].0.1 e+ = ec − F (xc )−1 F (xc ) = F (xc )−1 1 0 (F (xc ) − F (x∗ + tec ))ec dt. Then there is δ such that if x0 ∈ B(δ) the Newton iteration xn+1 = xn − F (xn )−1 F (xn ) converges q-quadratically to x∗ .1) We may also view x+ as the root of the two-term Taylor expansion or linear model of F about xc Mc (x) = F (xc ) + F (xc )(x − xc ). Then there are K > 0 and δ > 0 such that if xc ∈ B(δ) the Newton iterate from xc given by (5.1. This completes the proof of (5. Theorem 5.

we may terminate the iteration when the relative nonlinear residual F (x) / F (x0 ) is small. where δ is small enough so that the conclusions of Lemma 4.com/SIAM/FR16. (5. We have. a termination decision based on the relative residual may be made too late in the iteration or. the iteration may not terminate at all. We raised this issue in the context of linear equations in Chapter 1 when we compared (1.3. Since x0 ∈ B(δ) by assumption. .1 implies that (5. This electronic version is for personal use and may not be duplicated or distributed. If x ∈ B(δ). if there is error in the evaluation of F or the initial iterate is near a root. Then for all x ∈ B(δ) e F (x) 4κ(F (x∗ )) e ≤ ≤ . In all Buy this book from SIAM at http://www.1.3) then implies that xn → x∗ q-quadratically. Assume that the standard assumptions hold.1. Two examples are implicit integration of ordinary differential equations and differential algebraic equations. then if F (x∗ ) is well conditioned.1 twice.3.ec-securehost. Moreover. Termination of the iteration We can base our termination decision on Lemma 4.4). there is no “right-hand side” to use as a scaling factor and we must balance the relative and absolute size of the nonlinear residuals in some other way. where the initial iterate is an interpolation of a solution from a coarser computational mesh [99].1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. and solution of discretizations of partial differential equations. where the initial iterate is derived from the solution at the previous time step.1 hold.html. [83].3. The assumption that the initial iterate be “sufficiently near” the solution (x0 ∈ B(δ)) may seem artificial at first look. many situations in which the initial iterate is very near the root. 5. as we will see in the next section. 72 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Proof. [105]. [126]. Lemma 5. the entire sequence {xn } ⊂ B(δ). Let δ > 0 be small enough so that the conclusions of Lemma 4. ∗ )) 4 e0 κ(F (x F (x0 ) e0 where κ(F (x∗ )) = F (x∗ ) F (x∗ )−1 is the condition number of F (x∗ ) relative to the norm · .3) en+1 ≤ K en 2 ≤ η en < en and hence xn+1 ∈ B(ηδ) ⊂ B(δ). Let δ be small enough so that the conclusions of Theorem 5. however. However.1 hold for x ∈ B(δ). Theorem 5.2.1. the size of the relative nonlinear residual is a good indicator of the size of the error. as a direct consequence of applying Lemma 4.1 we conclude that if F (x∗ ) is well conditioned. In the nonlinear case.2) and (1. [16]. Then if n ≥ 0 and xn ∈ B(δ).3.2. local convergence results like those in this and the following two chapters describe the terminal phase of the iteration. Reduce δ if needed so that Kδ = η < 1. There are. a nonlinear form of Lemma 1. Therefore if xn ∈ B(δ) we may continue the iteration.1.2. From Lemma 5.1 hold. when the initial iterate is far from the root and methods such as those discussed in Chapter 8 are needed.1.

If one decides to continue. 5.Copyright ©1995 by the Society for Industrial and Applied Mathematics. whereas by terminating on small relative residuals or on a condition like (5. Hence the step is a reliable indicator of the error provided σ is not too near 1.1. say. [21].4) the decision to stop the iteration can be made before computation and factorization of F (xc ). the relation between the step and the error is less clear. and terminate the iteration when s is sufficiently small. In Chapter 6 we consider algorithms in which iterative methods are used to solve the equation for the Newton step. If computation and factorization of the Jacobian are inexpensive. However. [23]. in order to estimate ec this way. .4) F (x) ≤ τr F (x0 ) + τa . This electronic version is for personal use and may not be duplicated or distributed. Another way to decide whether to terminate is to look at the Newton step s = −F (xc )−1 F (xc ) = x+ − xc . one must do most of the work needed to compute x+ . make use of this in the special case of the chord method. Hence. Our examples and discussions of operation counts make the implicit assumption that the Jacobian is dense.html. and [151]. the issues for sparse Jacobians when direct methods are used to solve linear systems are very similar. Implementation of Newton’s method In this section we assume that direct methods will be used to solve the linear equation for the Newton step. Buy this book from SIAM at http://www. [23]. then termination on small steps becomes more attractive. NEWTON’S METHOD 73 of our MATLAB codes and algorithms for nonlinear equations our termination criterion is to stop the iteration if (5. where the relative error tolerance τr and absolute error tolerance τa are input to the algorithm.1.3. which implies that ec = s + O( ec 2 ). the Jacobian F (xc ) must be computed and factored. If the iteration is q-linearly convergent. Combinations of relative and absolute error tolerances are commonly used in numerical methods for ordinary differential equations or differential algebraic equations [16]. [21]. For methods other than Newton’s method. [151]. then e+ ≤ σ ec implies that (1 − σ) ec ≤ s ≤ (1 + σ) ec . However.com/SIAM/FR16. The differential equations codes discussed in [16]. This criterion is based on Theorem 5. In order to compute the Newton iterate x+ from a current point xc one must first evaluate F (xc ) and decide whether to terminate the iteration.ec-securehost. near the solution s and ec are essentially the same size.

Do while F (x) > τr r0 + τa (a) Compute F (x) (b) Factor F (x) = LU . Evaluation of F by finite differences should be expected to cost N times the cost of an evaluation of F because each column in F requires an evaluation of F to form the difference approximation. In many cases F can be computed more efficiently. chord(x. Buy this book from SIAM at http://www. The inputs are the same as for Newton’s method. newton(x. τa ) ∈ R2 . (c) Solve LU s = −F (x) (d) x = x + s (e) Evaluate F (x). automatic differentiation (see [94] for a collection of articles on this topic) may provide such an alternative. there may be little alternative to difference Jacobians. Compute F (x) 3.com/SIAM/FR16. 74 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Then the step is computed as the solution of F (xc )s = −F (xc ) and the iterate is updated x+ = xc + s. τ ) 1. and a vector of termination tolerances τ = (τr . accurately. In the future. r0 = F (x) 2.3. This means that the linear approximation of F (x) = 0 that is solved at each iteration has derivative determined by the initial iterate.2. x+ = xc − F (x0 )−1 F (xc ).ec-securehost.html.1. F. F. This electronic version is for personal use and may not be duplicated or distributed.3. For general nonlinearities or general purpose codes such as the MATLAB code nsol from the collection.7. Factor F (x) = LU . One approach to reducing the cost of items 2a and item 2b in the Newton iteration is to move them outside of the main loop. and directly than with differences and the analysis above for the cost of a Newton iterate is very pessimistic. Any other appropriate factorization such as QR or Cholesky could be used as well.Copyright ©1995 by the Society for Industrial and Applied Mathematics.21 for an example. the evaluation and factorization of F are the most costly. Factorization of F in the dense case costs O(N 3 ) floatingpoint operations. Algorithm 5. Of these steps. r0 = F (x) 2. See Exercise 5. the nonlinear map F . This method is called the chord method. Algorithm 5. The inputs of the algorithm are the initial iterate x. . For definiteness in the description of Algorithm newton we use an LU factorization of F . τ ) 1. Hence the cost of a Newton step may be roughly estimated as N + 1 evaluations of F and O(N 3 ) floating-point operations.

1. + (F (xc ) + ∆(xc ))−1 Buy this book from SIAM at http://www.html.e. We will refer to this phase of the iteration in which the nonlinear residual is no longer being reduced as stagnation of the iteration.. F (xc ) + ∆(xc ) is nonsingular) and satisfies (5. Let the standard assumptions hold.3. If ∆ is sufficiently small. Let δ be small enough so that the conclusions of Lemma 4.3. 5. there is no reason to expect that F (xn ) will ever be smaller than in general. This electronic version is for personal use and may not be duplicated or distributed. δ > 0.6) e+ ≤ K ec 2 + 2 F (xc )−1 − (F (xc ) + ∆(xc ))−1 ) F (x∗ ) ec (xc ) . Theorem 5. + Lemma 4.1. Proof. for example.1 imply (5. Let xN = xc − F (xc )−1 F (xc ) + and note that x+ = xN + (F (xc )−1 − (F (xc ) + ∆(xc ))−1 )F (xc ) − (F (xc ) + ∆(xc ))−1 (xc ).1 and Theorem 5. We continue in the next section with an analysis of the effects of errors in the function and Jacobian in a general context and then consider the chord method and difference approximation to F as applications.5) ¯ e+ ≤ K( ec 2 + ∆(xc ) ec + (xc ) ).4. Do while F (x) > τr r0 + τa (a) Solve LU s = −F (x) (b) x = x + s (c) Evaluate F (x). Errors in the function and derivative Suppose that F and F are computed inaccurately so that F + and F + ∆ are used instead of F and F in the iteration. the resulting iteration can return a result that is an O( ) accurate approximation to x∗ .com/SIAM/FR16.1 and Theorem 5. Similar differences arise if F (xc ) is numerically approximated by differences.Copyright ©1995 by the Society for Industrial and Applied Mathematics. If. The difference in the iteration itself is that an approximation to F (xc ) is used. .1 hold. These issues are discussed in [201] as well. Then there are ¯ K > 0.4. NEWTON’S METHOD 75 4. and δ1 > 0 such that if xc ∈ B(δ) and ∆(xc ) < δ1 then x+ = xc − (F (xc ) + ∆(xc ))−1 (F (xc ) + (xc )) is defined (i. This is different from convergence and was called “local improvement” in [65]. The only difference in implementation from Newton’s method is that the computation and factorization of the Jacobian are done before the iteration is begun.ec-securehost.1. is entirely due to floating-point roundoff.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

76 If

ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS

∆(xc ) ≤ F (x∗ )−1 ∆(xc ) ≤ F (xc )−1

−1

/4

then Lemma 4.3.1 implies that
−1

/2

and the Banach Lemma implies that F (xc ) + ∆(xc ) is nonsingular and (F (xc ) + ∆(xc ))−1 ≤ 2 F (xc )−1 ≤ 4 F (x∗ )−1 . Hence F (xc )−1 − (F (xc ) + ∆(xc ))−1 ≤ 8 F (x∗ )−1 We use these estimates and (5.6) to obtain e+ ≤ K ec Setting
2 2

∆(xc ) .

+ 16 F (x∗ )−1

2

F (x∗ ) ∆(xc ) ec + 4 F (x∗ )−1
2

(xc ) .

¯ K = K + 16 F (x∗ )−1

F (x∗ ) + 4 F (x∗ )−1

completes the proof. The remainder of this section is devoted to applications of Theorem 5.4.1. 5.4.1. The chord method. Recall that the chord method is given by

x+ = xc − F (x0 )−1 F (xc ). In the language of Theorem 5.4.1 (xc ) = 0, ∆(xc ) = F (x0 ) − F (xc ). Hence, if xc , x0 ∈ B(δ) ⊂ Ω (5.7) ∆(xc ) ≤ γ x0 − xc ≤ γ( e0 + ec ).

We apply Theorem 5.4.1 to obtain the following result. Theorem 5.4.2. Let the standard assumptions hold. Then there are KC > 0 and δ > 0 such that if x0 ∈ B(δ) the chord iterates converge q-linearly to x∗ and (5.8) en+1 ≤ KC e0 en . Proof. Let δ be small enough so that B(δ) ⊂ Ω and the conclusions of Theorem 5.4.1 hold. Assume that xn ∈ B(δ). Combining (5.7) and (5.5) implies ¯ ¯ en+1 ≤ K( en (1 + γ) + γ e0 ) en ≤ K(1 + 2γ)δ en . Hence if δ is small enough so that ¯ K(1 + 2γ)δ = η < 1

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

NEWTON’S METHOD

77

then the chord iterates converge q-linearly to x∗ . Q-linear convergence implies ¯ that en < e0 and hence (5.8) holds with KC = K(1 + 2γ). Another variation of the chord method is x+ = xc − A−1 F (xc ), where A ≈ F (x∗ ). Methods of this type may be viewed as preconditioned nonlinear Richardson iteration. Since ∆(xc ) = A − F (xc ) ≤ A − F (x∗ ) + F (x∗ ) − F (xc ) , if xc ∈ B(δ) ⊂ Ω then ∆(xc ) ≤ A − F (x∗ ) + γ ec ≤ A − F (x∗ ) + γδ. Theorem 5.4.3. Let the standard assumptions hold. Then there are KA > 0, δ > 0, and δ1 > 0, such that if x0 ∈ B(δ) and A − F (x∗ ) < δ1 then the iteration xn+1 = xn − A−1 F (xn ) converges q-linearly to x∗ and (5.9) en+1 ≤ KA ( e0 + A − F (x∗ ) ) en .

5.4.2. Approximate inversion of F . Another way to implement chordtype methods is to provide an approximate inverse of F . Here we replace F (x)−1 by B(x), where the action of B on a vector is less expensive to compute than a solution using the LU factorization. Rather than express the iteration in terms of B −1 (x) − F (x) and using Theorem 5.4.3 one can proceed directly from the definition of approximate inverse. Note that if B is constant (independent of x) then the iteration x+ = xc − BF (xc ) can be viewed as a preconditioned nonlinear Richardson iteration. We have the following result. Theorem 5.4.4. Let the standard assumptions hold. Then there are KB > 0, δ > 0, and δ1 > 0, such that if x0 ∈ B(δ) and the matrix-valued function B(x) satisfies (5.10) I − B(x)F (x∗ ) = ρ(x) < δ1

for all x ∈ B(δ) then the iteration xn+1 = xn − B(xn )F (xn ) converges q-linearly to x∗ and (5.11) en+1 ≤ KB (ρ(xn ) + en ) en .

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

78

ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS

Proof. On one hand, this theorem could be regarded as a corollary of the Banach Lemma and Theorem 5.4.1. We give a direct proof. First, by (5.10) we have (5.12) B(x) = B(x)F (x∗ )F (x∗ )−1 ≤ B(x)F (x∗ ) F (x∗ )−1 ≤ MB = (1 + δ1 ) F (x∗ )−1 .

Using (5.12) and e+ = ec − B(xc )F (xc ) =
1 0

(I − B(xc )F (x∗ + tec ))ec dt
1 0

= (I − B(xc )F (x∗ ))ec + B(xc ) we have

(F (x∗ ) − F (x∗ + tec ))ec dt

e+ ≤ ρ(xc ) ec + MB γ ec 2 /2.

This completes the proof with KB = 1 + MB γ/2. 5.4.3. The Shamanskii method. Alternation of a Newton step with a sequence of chord steps leads to a class of high-order methods, that is, methods that converge q-superlinearly with q-order larger than 2. We follow [18] and name this method for Shamanskii [174], who considered the finite difference Jacobian formulation of the method from the point of view of efficiency. The method itself was analyzed in [190]. Other high-order methods, especially for problems in one dimension, are described in [190] and [146]. We can describe the transition from xc to x+ by y1 (5.13) = xc − F (xc )−1 F (xc ),

yj+1 = yj − F (xc )−1 F (yj ) for 1 ≤ j ≤ m − 1, x+ = ym .

Note that m = 1 is Newton’s method and m = ∞ is the chord method with {yj } playing the role of the chord iterates. Algorithmically a second loop surrounds the factor/solve step. The inputs for the Shamanskii method are the same as for Newton’s method or the chord method except for the addition of the parameter m. Note that we can overwrite xc with the sequence of yj ’s. Note that we apply the termination test after computation of each yj . Algorithm 5.4.1. sham(x, F, τ, m) 1. r0 = F (x) 2. Do while F (x) > τr r0 + τa

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

html.2 hold.4. In fact. The analysis in [18] and the expectation that. Assume that we compute F (x) + (x) instead of F (x) and attempt to approximate the action of F on a vector by a forward difference.2. if we set y1 = xn . ≤ KC en j+1 . NEWTON’S METHOD 79 (a) Compute F (x) (b) Factor F (x) = LU .14) en+1 ≤ KS en m+1 . (5. Then there are KS > 0 and δ > 0 such that if x0 ∈ B(δ) the Shamanskii iterates converge q-superlinearly to x∗ with q-order m + 1 and (5. for example. .4.4. Evaluate F (x).8) implies that for 1≤j≤m yj −x∗ ≤ KC xn −x∗ yj−1 −x∗ = KC en j yj−1 −x∗ ≤ . Let the standard assumptions hold and let m ≥ 1 be given. indicate that the chord method is usually the best option for very large problems. Algorithm nsol is based on this idea and using an idea from [114] computes and factors the Jacobian only until an estimate of the q-factor for the linear convergence rate is sufficiently low. Setting j = m in the above inequality completes the proof. Proof. . Theorem 5. . if the initial iterate is sufficiently accurate. m i. If F (x) ≤ τr r0 + τa exit. The convergence result is a simple application of Theorem 5. Optimal choices of m as a function of N can be derived under assumptions consistent with dense matrix algebra [18] by balancing the O(N 3 ) cost of a matrix factorization with the cost of function and Jacobian evaluation. The advantage of the Shamanskii method over Newton’s method is that high q-orders can be obtained with far fewer Jacobian evaluations or factorizations.5. Solve LU s = −F (x) ii. h Buy this book from SIAM at http://www. Here the jth column of F (x) is F (x)ej where ej is the unit vector with jth component 1 and other components 0. .4. A forward difference approximation to F (x)w would be F (x + hw) + (x + hw) − F (x) − (x) .2. . Hence xn+1 ∈ B(δ). . iv.Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed. x = x + s iii.4. Difference approximation to F . when building an approximate Jacobian.ec-securehost. only a small number of iterations will be taken.4.com/SIAM/FR16. Let δ be small enough so that B(δ) ⊂ Ω and that the conclusions of Theorem 5. We would do this. (c) for j = 1. Then if xn ∈ B(δ) all the intermediate iterates {yj } are in B(δ) by Theorem 5. 5.

the cost of a forward difference approximation of F (x)w is an additional evaluation of F (at the point x+hw). (∇h F )(x) is the N × N matrix whose jth column is given by   F (x + h x ej ) − F (x)     h x   F (he ) − F (x)   j  x=0 x=0 (∇h F )(x)j = h We make a similar definition of the difference approximation of the directional derivative. that very small difference steps h can lead to inaccurate results. A more important consequence of this analysis is that the choice of step size in a forward difference approximation to the action of the Jacobian on a vector must take into account the error in the evaluation of F . The reason for this is that the derivative has been approximated. h should be scaled to reflect that. .Copyright ©1995 by the Society for Industrial and Applied Mathematics. The choice h ≈ 10−15 in this case can lead to disaster (try it!). Definition 5.ec-securehost. Because of this we must carefully specify what we mean by a difference approximation to the Jacobian.2.1 below specifies the usual choice.html. Definition 5. Exploitation of special structure could reduce this cost.com/SIAM/FR16. for instance. Definition 5. Let F be defined in a neighborhood of x ∈ RN and let Buy this book from SIAM at http://www. This means. If this is not the case. the analysis here indicates that h ≈ 10−7 is a reasonable choice. This can become very important if part of F is computed from measured data.4. h F (x)w − √ The quantity inside the O-term is minimized when h = ¯. Let F be defined in a neighborhood of x ∈ RN . Here we have made the implicit assumption that x and w are of roughly the same size.4.4. 80 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Assume that (x) ≤ ¯. and so linearity has been lost. The forward difference approximation to the directional derivative is not a linear operator in w. one for each column of the Jacobian. In the special case where (x) is a result of floating-point roundoff in full precision (¯ ≈ 10−15 ). If F (x) has already been computed. The choice h = ¯1/2 x / w reflects all the discussion above. Then F (x + hw) + (x + hw) − F (x) − (x) = O(h + ¯/h). This electronic version is for personal use and may not be duplicated or distributed. Hence the evaluation of a full finite difference Jacobian would cost N function evaluations.1.

If information on the sparsity pattern of the Jacobian is available it is sometimes possible [47]. Hence one will see quadratic convergence until the error in F admits no further reduction. We have (5.4.1. Theorem 5. Then there are δ. and (x) ≤ ¯ for all x ∈ B(δ). When en ≤ ¯.4.com/SIAM/FR16. ¯ Lemma 5.6. there is no guarantee that (xn ) → 0 as xn → x∗ . This electronic version is for personal use and may not be duplicated or distributed. Let the standard assumptions hold. to evaluate ∇h F with far fewer than N evaluations of F . one should expect the sequence { en } to stagnate and cease to decrease once√ en ≈ ¯. while en >> ¯.4. Therefore in the early phases of the iteration.1 implies the following result. then ∆(x) = O( ¯) by Lemma 5.ec-securehost.1.4. the progress of the iteration will be indistinguishable from Newton’s method as implemented with an exact √ Jacobian. [42]. h > ¯ 0 such that if x ∈ B(δ). . x = 0. Hence. Theorem 5.4.6 does not imply that an iteration will converge to Even if is entirely due to floating-point roundoff. NEWTON’S METHOD 81 w ∈ RN . x+ = xc − ∇hc (F + )(xc )−1 (F (xc ) + (xc )) Buy this book from SIAM at http://www.html. w = 0. and there is M− > 0 such that √ hc > M− ¯ then satisfies e+ ≤ KD (¯ + ( ec + hc ) ec ). h ≤ h. w. h x = 0. (x) ≤ ¯ for all x ∈ B(δ).15)Dh F (x : w) =   0.4.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Note that Theorem 5. w w The standard assumptions and the Banach Lemma imply the following result. Let the standard assumptions hold.6 says that en+1 = O(¯). A difference approximation to a full Jacobian does not use any special structure of the Jacobian and will probably be less efficient than a hand coded analytic Jacobian. h x F (hw/ w ) − F (x) . Then there are positive δ. ¯. F (x + h x w/ w ) − F (x) . Theorem 5. If we assume that the forward difference step has been selected so that √ √ h = O( ¯). and KD such that if xc ∈ B(δ). then ∇h (F + )(x) is nonsingular for all x ∈ B(δ) and there is MF > 0 such that F (x) − ∇h (F + )(x) ≤ MF (h + ¯/h).                    w = 0. ¯. If the sparsity pattern is such that F can be reduced to x∗ .

an = xn − xn−1 The resulting method is called the secant method and is due to Newton [140].4.Copyright ©1995 by the Society for Industrial and Applied Mathematics. While we used full matrix approximations to numerical Jacobians in all the examples reported here.7. Then there is δ > 0 such that if x0 .4. (x) = 0.1. Let the standard assumptions hold with N = 1. . If we set s = xn − xn−1 we have by (5. There is no reason to use a difference approximation to f for equations in a single variable because one can use previously computed data to approximate f (xn ) by f (xn ) − f (xn−1 ) . 82 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS a block triangular form. we are done. The standard assumptions. 5. If xn = x∗ for some finite n. a linear function of w and it is usually the case that (∇h F (x))w = Dh F (x : w). imply that f (x∗ ) = 0. x0 . which uses the size of w.4. one needs two approximations to x∗ . the error in the difference approximation will be large. Theorem 5.e. [137].4. Let xc be the current iterate and x− the previous iterate.16) and the fundamental theorem of calculus ∆(xn ) = 1 0 (f (xn−1 + ts) − f (xn )) dt Buy this book from SIAM at http://www. Moreover. We will prove that the secant method is locally q-superlinearly convergent if the standard assumptions hold. The local convergence theory can be easily described by Theorem 5. The results in this section are for single equations f (x) = 0. We assume for the sake of simplicity that there are no errors in the evaluation of f . the entire solution process can take advantage of this structure [59]. Assume that x−1 . For problems in which the Lipschitz constant of F is large. xn ∈ B(δ). The secant method.. Proof.5. xc − x− We have the following theorem. . they should be used with caution. x0 and x−1 . the difference between them may be significant and a difference approximation to a directional derivative.com/SIAM/FR16. too. Dh F (x : w) is not. Let δ be small enough so that B(δ) ⊂ Ω and the conclusions of Theorem 5. In order to start. is likely to be more accurate than ∇h F (x)w. If F (x∗ ) is ill conditioned or the Lipschitz constant of F is large. We have (xc ) = 0 and (5.html. where f is a real-valued function of one real variable. . . This electronic version is for personal use and may not be duplicated or distributed.1 hold. i.16) ∆(xc ) = f (xc ) − f (x− ) − f (xc ). in general. then. x−1 ∈ B(δ) and x0 = x−1 then the secant iterates converge q-superlinearly to x∗ . Otherwise xn = xn−1 . .ec-securehost.

r− = βγ We will let B0 be the closed ball of radius r− about x0 B0 = {x | x − x0 ≤ r− }. However.1. The secant method and the classical bisection (see Exercise 5. 2 and hence |∆(xn )| ≤ γ|xn−1 − xn |/2 ≤ Applying Theorem 5. which states that if the standard assumptions “almost” hold at a point x0 . This completes the proof. and F (x0 )−1 F (x0 ) ≤ η. 2.8) method are the basis for the popular method of Brent [17] for the solution of single equations.17) ¯ |en+1 | ≤ K((1 + γ/2)|en |2 + γ|en ||en−1 |/2).Copyright ©1995 by the Society for Industrial and Applied Mathematics.com/SIAM/FR16. F (x0 )−1 ≤ β. and γ with βηγ ≤ 1/2 and x0 ∈ RN ¯ such that 1. We do not prove the result in its full generality and refer to [106]. There are β.html.5. This result is of use.ec-securehost. We require the following assumption. then there is a root and the Newton iteration converges r-quadratically to that root. 5. . F is differentiable at x0 . The Kantorovich Theorem In this section we present the result of Kantorovich [107].7. r. we give a simplified version of a Kantorovich-like result for the chord Buy this book from SIAM at http://www. Therefore |en+1 | ¯ ≤ K((1 + γ/2)|en | + γ|en−1 |/2) → 0 |en | as n → ∞. [106]. NEWTON’S METHOD 83 γ(|en | + |en−1 |) . F is Lipschitz continuous with Lipschitz constant γ in a ball of radius r ≥ r− about x0 where ¯ √ 1 − 1 − 2βηγ . in proving that discretizations of nonlinear differential and integral equations have solutions that are near to that of the continuous problem. for example. η. This electronic version is for personal use and may not be duplicated or distributed.5. ¯ K(1 + γ)δ = η < 1 If we reduce δ so that then xn → x∗ q-linearly. and [58] for a complete proof and for discussions of related results. [145].4. Assumption 5. [57].1 implies (5.

5. Buy this book from SIAM at http://www. βγ about x0 . Theorem 5. Let Assumption 5. and xn ∈ B0 for all n.1 hold and let (5. 2 φ(x) − x0 ≤ η + βγr− /2 = r− .html.1 we have for all x ∈ S F (x0 )−1 F (x) = F (x0 )−1 F (x0 ) +F (x0 )−1 (5. the chord iteration with x0 as the initial iterate converges to x∗ q-linearly with q-factor βγr− . Proof.1 assumes a bit more than Assumption 5. and [57]. Theorem 5. This electronic version is for personal use and may not be duplicated or distributed.1 but has a simple proof that uses only the contraction mapping theorem and the fundamental theorem of calculus and obtains q-linear convergence instead of r-linear convergence. By the fundamental theorem of calculus and Assumption 5.ec-securehost. Moreover x∗ is the unique root of F in the ball of radius √ 1 + 1 − 2βηγ ¯ min r.5. We will show that the map φ(x) = x − F (x0 )−1 F (x) is a contraction on the set S = {x | x − x0 ≤ r− }. Unlike the chord method results in [106].Copyright ©1995 by the Society for Industrial and Applied Mathematics. 84 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS method to show how existence of a solution can be derived from estimates on the function and Jacobian.com/SIAM/FR16. Then there is a unique root x∗ of F in B0 .18) βγη < 1/2. 1 0 (F (x0 + t(x − x0 )) − F (x0 ))(x − x0 ) dt. .5.1.19) 1 0 F (x0 + t(x − x0 ))(x − x0 ) dt = F (x0 )−1 F (x0 ) + x − x0 +F (x0 )−1 1 0 (F (x0 + t(x − x0 )) − F (x0 ))(x − x0 ) dt and φ(x) − x0 = −F (x0 )−1 F (x0 ) −F (x0 )−1 Hence.5.5. [145].

ec-securehost. ≤ 2n βγ n Buy this book from SIAM at http://www. Clearly x∗ is a root of F because it is a fixed point of φ. then we are done by the uniqueness of x∗ in B0 . Hence we may assume that r > r− .19) F (x0 )−1 F (x) ≥ r − η − βγr2 /2 > 0 because r− < r < r+ .5.1 and has a very different proof. βγ about x0 and the errors satisfy (5. Theorem 5. Letting r = x − x0 we have by (5. y ∈ S.2.20) en (2βηγ)2 .5. the Newton iteration with x0 as the initial iterate converges to x∗ . If r = r− .Copyright ©1995 by the Society for Industrial and Applied Mathematics. for all x.1 hold. 1 0 Therefore. If we do this. We state the theorem in the standard way [145]. This electronic version is for personal use and may not be duplicated or distributed. and xn ∈ B0 for all n. Note that for all x ∈ B0 φ (x) = I − F (x0 )−1 F (x) = F (x0 )−1 (F (x0 ) − F (x)) and hence φ (x) ≤ βγ x − x0 ≤ βγr− < 1. Let Assumption 5.com/SIAM/FR16.html. φ(x) − φ(y) = φ (y + t(x − y))(x − y) dt ≤ βγr− x − y . [106]. This proves the convergence result and the uniqueness of x∗ in B0 . [63]. The Kantorovich theorem is more precise than Theorem 5.5.2. To show that φ is a contraction on S and to prove the q-linear convergence we will show that φ(x) − φ(y) ≤ βγr− x − y for all x. x∗ is the unique root of F in the ball of radius √ 1 + 1 − 2βηγ ¯ min r. r+ ) r then F (x) = 0. This completes the proof. To prove the remainder of the uniqueness assertion let √ 1 + 1 − 2βηγ r+ = βγ ¯ and note that r+ > r− because βηγ < 1/2. We must show ¯ that if x is such that r− < x − x0 < min(¯.1) will imply that there is a unique fixed point x∗ of φ in S and that xn → x∗ q-linearly with q-factor βγr− . y ∈ B0 . NEWTON’S METHOD 85 and φ maps S into itself. Then there is a unique root x∗ of F in B0 . . the contraction mapping theorem (Theorem 4. We know that βγr− < 1 by our assumption that βηγ < 1/2. using the r-quadratic estimates from [106].

τ. or a combination of all of them based on the reduction in the nonlinear residual. σ = 0 (d) Do While is < m and σ ≤ ρ i. x = x + s iv. If F (x+ ) / F (xc ) is below a given threshold. In addition to the inputs of Algorithm sham.6.Copyright ©1995 by the Society for Industrial and Applied Mathematics. F. the Shamanskii method.7. nsol(x.1. Do while F (x) > τr r0 + τa (a) Compute ∇h F (x) ≈ F (x) (b) Factor ∇h F (x) = LU . The output of the MATLAB implementation of nsol is the solution. If F (x) ≤ τr r0 + τa exit.20) implies that the convergence is r-quadratic (see Exercise 5. the chord method. which is not easy.25. Evaluate F (x) v. however. If the ratio is too large. Algorithm 5. ρ) 1.html. we compute and factor F (x+ ) for use in the subsequent chord steps. Examples for Newton’s method In the collection of MATLAB codes we provide an implementation nsol of Newton’s method. . m.ec-securehost. This indicates that the local convergence analysis in this Buy this book from SIAM at http://www. The Jacobian is recomputed and factored if either the ratio of successive residuals exceeds the threshold 0 < ρ < 1 or the number of iterations without an update exceeds m. 5.6.com/SIAM/FR16. rc = r+ vi. σ = r+ /rc . the ideas in the implementation of nsol also apply to situations in which the Jacobian is sparse and direct methods for solving the equation for the Newton step are necessary. This electronic version is for personal use and may not be duplicated or distributed. and an error flag.16). the reader is asked to create a sparse form of nsol. the Jacobian is not recomputed. In Exercise 5. Solve LU s = −F (x) iii. One would still want to minimize the number of Jacobian computations (by the methods of [47] or [42]. (c) is = 0. This latter option decides if the Jacobian should be recomputed based on the ratio of successive residuals. r+ = F (x) .7. the history of the iteration stored as the vector of l∞ norms of the nonlinear residuals. say) and sparse matrix factorizations. nsol uses diffjac to approximate the Jacobian by ∇h F with h = 10−7 . 86 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS If βηγ < 1/2 then (5. In the MATLAB code the iteration is terminated with an error condition if σ ≥ 1 at any point. While our examples have dense Jacobians. is = is + 1 ii. a threshold for the ratio of successive residuals is also input. rc = r0 = F (x) 2.

9. We used the function identically one as initial iterate. nsol allows the user to specify a maximum number of iterations.com/SIAM/FR16.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Many times the Jacobian can be computed efficiently by reusing results that are already available from the function evaluation. nsol is based on the assumption that a difference Jacobian will be used. [30]. The standard assumptions hold near either solution [110] for c ∈ [0. where µi = (i − 1/2)/N for 1 ≤ i ≤ N . Because H has a singularity at µ = 0. (5. will find the physically meaningful solution [110].5.21) and the discrete analog (5. The resulting discrete problem is  (5. It is known [132] that both (5. and the ratio Rn = F (xn ) ∞ / F (xn−1 ) ∞ Buy this book from SIAM at http://www. F (xn ) ∞ / F (x0 ) ∞ .22) have two solutions for c ∈ (0.21) F (H)(µ) = H(µ) − 1 − The Chandrasekhar H-equation. The Chandrasekhar H-equation. With these default parameters the decision to update the Jacobian is made based entirely on the reduction in the norm of the nonlinear residual. NEWTON’S METHOD 87 section is not applicable. We will discretize the equation with the composite midpoint rule. Only one of these solutions has physical meaning. is used to solve exit distribution problems in radiative transfer.html. The parameters m and ρ are assigned the default values of 1000 and . In Chapter 8 we discuss how to recover from this. the solution of the discrete problem is not even a first-order accurate approximation to that of the continuous problem. We set the termination parameters τr = τa = 10−6 . Here we approximate integrals on [0. In Exercise 5. however.22) F (x)i = xi − 1 − c 2N µi xj  µi + µj j=1 N −1 .21 you are asked to modify the code to accept an analytic Jacobian. This electronic version is for personal use and may not be duplicated or distributed. with the 0 function or the constant function with value 1 as initial iterate. 1] by 1 0 f (µ) dµ ≈ 1 N N j=1 f (µj ). .1 we tabulate the iteration counter n. We computed all Jacobians with the MATLAB code diffjac.7. We present some iteration statistics in both tabular and graphical form. In the computations reported here we set N = 100 and c = . 1). 1). The discrete problem has its own physical meaning [41] and we will attempt to solve it to high accuracy. We begin with a comparison of Newton’s method and the chord method.ec-securehost. the default is 40. Newton’s method. In Table 5. c 2 1 µH(ν) dν 0 −1 µ+ν = 0. [41].

729e-07 ∞ Rn 1. for this problem. the cost of a single evaluation and factorization of the Jacobian is far more than the cost of the entire remainder of the chord iteration. 5.7.3 million.7. the iteration statistics.136e-01 2. In Fig. The concave iteration track is the signature of superlinear convergence.000e+00 1. In Fig.21. The reader is encouraged to do this in Exercise 5. 88 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Table 5.136e-01 for n ≥ 1 for both Newton’s method and the chord method.000e+00 1.136e-01 2. as compared with over 8. The Newton iteration required over 8.7.074e-02 6.html.1 the solid curve is a plot of F (xn ) ∞ / F (x0 ) ∞ for Newton’s method and the dashed curve a plot of F (xn ) ∞ / F (x0 ) ∞ for the chord method.823e-02 2.865e-04 F (xn ) ∞ / F (x0 ) 1.15 for a more careful examination of this concavity. the cost of the Jacobian computation can be reduced by analytic evaluation and reuse of data that have already been computed for the evaluation of F . . In the context of this particular example.480e-01 1. Buy this book from SIAM at http://www.ec-securehost.077e-01 2. we plot. In the case considered here. The good performance of the chord method is no accident. 5. In Exercise 5.891e-06 ∞ Rn 1. n 0 1 2 3 4 5 6 7 8 F (xn ) ∞ / F (x0 ) 1. In the remaining examples.511e-03 1.8 for Newton’s method. See Exercise 5.132e-01 2.1 Comparison of Newton and chord iteration.com/SIAM/FR16.Copyright ©1995 by the Society for Industrial and Applied Mathematics. The linear convergence of the chord method can also be seen in the convergence of the ratios Rn to a constant value. but do not tabulate. For this example the MATLAB flops command returns an fairly accurate indicator of the cost. the linear track indicates linear convergence. This electronic version is for personal use and may not be duplicated or distributed.118e-01 2.480e-01 2.334e-05 1. If we think of the chord method as a preconditioned nonlinear Richardson iteration with the initial Jacobian playing the role of preconditioner.2 we compare the Shamanskii method with m = 2 with Newton’s method. then the chord method should be much more efficient than Newton’s method if the equations for the steps can be solved cheaply.698e-03 7. The Shamanskii computation required under 6 million floating-point operations.480e-01 3.965e-04 6. the simple chord method is the most efficient. However.353e-05 2.388e-03 2.18 you are asked to obtain timings in your environment and will see how the floating-point operation counts are related to the actual timings.8 million floatingpoint operations while the chord iteration required under 3. The ideas in Chapters 6 and 7 show how the low iteration cost of the chord method can be combined with the small number of iterations of a superlinearly convergent nonlinear iterative method.136e-01 2.480e-01 2.

The q-factor for the chord method in this computation was over .5 3 3.com/SIAM/FR16. 5. the chord method requires 188 iterations and 10.6 million floating-point operations.2. For this problem the Jacobian at the solution is nearly singular [110]. We now reconsider the H-equation with c = .html. Aside from c.9. 5. This electronic version is for personal use and may not be duplicated or distributed. Newton and chord methods for H-equation.96.Copyright ©1995 by the Society for Industrial and Applied Mathematics. c = .9. The hybrid approach in nsol with the default parameters would be the chord method for this problem. 10 0 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 10 -6 10 -7 0 0.1. NEWTON’S METHOD 0 89 10 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 10 -6 10 -7 0 1 2 3 4 Iterations 5 6 7 8 Fig. Newton and Shamanskii method for H-equation.9999.5 2 Iterations 2. c = . The hybrid method in nsol with the default parameters Buy this book from SIAM at http://www. While Newton’s method converges in 7 iterations and requires 21 million floating-point operations.ec-securehost. indicating very slow convergence.5 4 Fig. .5 1 1. all iteration parameters are the same as in the previous example.

in general. converges in 14 iterations. This electronic version is for personal use and may not be duplicated or distributed. In practice. the theory does not. however. The theory for the chord method asserts that the error norms converge q-linearly to zero. 90 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS computes the Jacobian four times during the course of the iteration. 5. One should approach plots of convergence histories with some caution.com/SIAM/FR16. While the plots indicate q-linear convergence of the nonlinear residuals.3. In Fig. This implies only that the nonlinear residual norms { F (xn ) } converge r-linearly to zero.9999. 10 0 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 10 -6 0 2 4 6 8 Iterations 10 12 14 Fig. 5. .3 we compare the Newton iteration with the hybrid.html. Buy this book from SIAM at http://www.Copyright ©1995 by the Society for Industrial and Applied Mathematics. plots like those in this section are common.ec-securehost. support this. Newton and hybrid methods for H-equation c = . and requires 12 million floating-point operations.

2.7. Prove that the secant method converges r-superlinearly with r-order √ ( 5 + 1)/2. When asked to do this. If f (y) = 0. Given the standard assumptions. and in the rest of this part of the book. xk ) that r l contain a root of a function of one variable.7. for each iteration tabulate the iteration counter. xk ) with r l f (xk )f (xk ) < 0 we replace one of the endpoints with y = (xk + xk )/2 r r l l to force f (xk+1 )f (xk+1 ) < 0. A better alternative in the MATLAB environment is to use the semilogy command to plot the norms.ec-securehost. Suppose F (xn ) is replaced by ∇hn F (xn ) in the Shamanskii method.1. Show that the secant method converges q-superlinearly with q-order √ ( 5 + 1)/2. x0 ) such that f (x0 )f (x0 ) < 0.html. 5. you will be asked to plot or tabulate iteration statistics. 5. for nonsingular N × N matrices A and B. 5. can the Newton iterates for F (x) = 0 and AF (Bx) = 0 show any performance difference when started at the same initial iterate? What about the chord method? 5.7. r r l l Buy this book from SIAM at http://www.7. Show that if the Lipschitz G(x) − G(y) ≤ K x − y continuity condition on F in the standard assumptions is replaced by H¨lder continuity with exponent α > 0 that the Newton iterates converge o with q-order 1 + α.7. This is easy.7. Use the l∞ norm. we terminate the iteration. The bisection method produces a sequence of intervals (xk .5.4.1.4. 5.Copyright ©1995 by the Society for Industrial and Applied Mathematics.7. r l Show that the sequence xk = (xk + xk )/2 is r-linearly convergent if f is r l continuous and there exists (x0 . Let ∇h F be given by Definition 5. Discuss how hn must be decreased as the iteration progresses to preserve the q-order of m + 1 [174]. This electronic version is for personal use and may not be duplicated or distributed.6. Can the performance of the Newton iteration be improved by a linear change of variables? That is. A function G is said to be H¨lder continuous with exponent α in Ω if o α for all x. Exercises on Newton’s method In some of the exercises here. When is it locally qquadratically convergent? 5. . NEWTON’S METHOD 91 5.com/SIAM/FR16. 5. one can visually inspect the plot to determine superlinear convergence (concavity) without explicitly computing the ratios. the norm of F . prove that the iteration given by xn+1 = xn − (∇hn F (xn ))−1 F (xn ) is locally q-superlinearly convergent if hn → 0. In what way is the convergence analysis of the secant method changed if there are errors in the evaluation of f ? 5. and the ratio of F (xn ) / F (xn−1 ) for n ≥ 1.3.7.7. y ∈ Ω.8.7. When one does this. Given (xk .

15.Copyright ©1995 by the Society for Industrial and Applied Mathematics.99x0 . Show that if f is a real function of one real variable. 5. and the secant method. and f (x∗ ) = f (x∗ ) = 0 but f (x∗ ) = 0 then the iteration xn+1 = xn − 2f (xn )/f (xn ) converges locally q-quadratically to x∗ provided x0 is sufficiently near to x∗ .7. but not equal to x∗ [171]. f (x∗ ) = 0 and f (x) is Lipschitz continuous.11. . For scalar functions of one variable a quadratic model of f about a point xc is mc (x) = f (xc ) + f (xc )(x − xc ) + f (xc )(x − xc )2 /2. Can the ideas in [170] be applied to some of the high-order methods discussed in [190] for equations in one unknown? 5. Test it on the problems in Exercise 5. . and explain your results. In what cases is this algorithm superior to Newton’s method? For hints and generalizations of this idea to functions of several variables. 5.14. x0 = . f (x) = x2 + 1. .7. 5. Show that if f is a real function of one real variable.5 2. f (x) = sin(x). This electronic version is for personal use and may not be duplicated or distributed. Write a program that solves single nonlinear equations with Newton’s method. 5.5 5.13. x0 is sufficiently near x∗ . x0 = 3 4. Apply your program to the following function/initial iterate combinations. f (x) = x2 . f (x) = cos(x) − x. and f (x0 )f (x0 ) > 0. the standard assumptions hold.html. What if f (x∗ ) = 0? How can you extend this result if f (x∗ ) = f (x∗ ) = . How would you estimate the q-order of a q-superlinearly convergent sequence? Rigorously justify your answer.7. If xn → x∗ q-superlinearly with q-order α > 1.7. show that log( en ) is a concave function of n in the sense that log( en+1 ) − log( en ) Buy this book from SIAM at http://www. f (x) = tan−1 (x). f is Lipschitz continuous.7. then the Newton iteration converges monotonically to x∗ . What problems might arise in such a method? Resolve these problems and implement your algorithm.com/SIAM/FR16. the chord method. tabulate or plot iteration statistics. f (k) (x∗ ) = 0 = f (k+1) (x∗ )? See [131]. 92 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 5.9. x0 = . x0 = 10 5.9. then the Newton iteration converges cubically (q-superlinear with q-order 3).ec-securehost. see [170]. x0 = 1 3.10.7. 1. Consider the iterative method that defines x+ as a root of mc . Show that if f is a real function of one real variable.7.7. the standard assumptions hold.12. For the secant method. use x−1 = .

[184]) the computed LU factorization of F (x) is the exact LU factorization of a nearby matrix. Compare timings and flop counts with Newton’s method. [75].9 (see [52]). and the algorithm in nsol. Duplicate the results on the H-equation in § 5.18. Is this still the case if the convergence is q-superlinear but there is no q-order? What if the convergence is q-linear? 5.7. the discretized nonlinearity from the H equation in (5. Tabulate or plot iteration statistics.9999 with fixed-point iteration.20. Typically (see [89]. or [118] for acceleration of the convergence. Modify dirder by setting the forward difference step h = 10−2 . How does the analysis for the QR factorization differ? 5. Solve the H-equation with c = 1 and N = 100. Does the trick in Exercise 5. Use the MATLAB cputime command to see how long the entire iteration took for each N. the chord method.7.20) is r-quadratically (but not q-quadratically) convergent to 0 if βηγ < 1/2. Modify nsol to accept this analytic Jacobian rather than evaluate a difference Jacobian. Solve the H-equation with N = 100.4. Compute by hand the Jacobian of F . 5. How does this change affect the convergence rate observations that you have made? Buy this book from SIAM at http://www. [49].7. Prove that F can be computed at a cost of less than twice that of an evaluation of F . This electronic version is for personal use and may not be duplicated or distributed.1 to describe how this factorization error affects the convergence of the Newton iteration.16.7. Compare execution times in your environment.9 and c = .17. Show that the sequence on the right side of (5. How does the performance of the methods differ as N and c change? Compare your results to the tabulated values of the H function in [41] or [14]. [170].6. Use the theory in [89]. 5. Explain your results (see [50] and [157]).html. [184] or another book on numerical linear algebra and Theorem 5. Try the chord method and compare the results to part 4 of Exercise 5.7. c combination and the MATLAB flops command to get an idea for the number of floatingpoint operations required by the various methods.11 improve the convergence? Try one of the approaches in [51]. [35] and [155] for more discussion of this.22). Try different values of N and c. LU = F (x) + E. 5. How do the execution times and flop counts change? 5. 5.Copyright ©1995 by the Society for Industrial and Applied Mathematics.7.7. .7.ec-securehost. c = .21.com/SIAM/FR16.19.7.22. You might also look at [90]. NEWTON’S METHOD 93 is a decreasing function of n for n sufficiently large.

the solution is u∗ (x) = x(1 − x). What about the chord method? 5.23) F (φ. λ) of an N ×N matrix A by solving the system of N + 1 nonlinear equations (5.7. you could modify diffjac to use the techniques of [47] and nsol to use the sparse matrix factorization in MATLAB.html. Add noise to your function evaluation with a random number generator. Develop a sparse version of nsol for problems with tridiagonal Jacobian. λ)T ∈ RN +1 . and f (x) = 2 − sin(x(1 − x)).7.ec-securehost. In MATLAB.7. Suppose you try to find an eigenvector-eigenvalue pair (φ. λ) ∈ RN +1 is an eigenvector-eigenvalue pair. . λ) = Aφ − λφ φT φ − 1 = 0 0 for the vector x = (φT . Here. how are the convergence rates different? What is an appropriate termination criterion? At what point does the iteration stagnate? 5. This electronic version is for personal use and may not be duplicated or distributed. when is F (x) nonsingular? Relate the application of Newton’s method to (5. Apply your code to the central difference discretization of the two-point boundary value problem −u = sin(u) + f (x). If the size of the noise is O(10−4 ) and you make the appropriate adjustments to dirder.24. What is the Jacobian of this system? If x = (φ. Buy this book from SIAM at http://www.7. for example.25. See [150] for more development of this idea.23.23) to the inverse power method. 94 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 5. such as the the MATLAB rand function. the cost of a Jacobian is O(N ) function evaluations. and that x0 is near enough to x∗ so that the Newton iteration converges q-quadratically to x∗ .com/SIAM/FR16. Assume that the standard assumptions hold.26.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Show that the number n of Newton iterations needed to obtain en ≤ e0 is O(log( )) and that the number of floating-point operations required is O(N 3 log( )). u(0) = u(1) = 0. that the cost of a function evaluation is O(N 2 ) floating-point operations. 5. Let u0 = 0 be the initial iterate.

1.1 and the other results in Chapter 5.2 we show how the requirements of the simple analysis can be changed to admit a more aggressive (i. for discussion of these ideas in the context of the classical stationary iterative methods.1 we give a straightforward analysis that uses the techniques of Chapter 5.e. This is quite useful because conditions like (6. In § 6.1) are precisely the small linear residual termination conditions for iterative solution of the linear system for the Newton step.1) F (xc )s + F (xc ) ≤ ηc F (xc ) are considered.ec-securehost.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Before that we will give the basic result from [55] to show how a step that satisfies (6. but the other Krylov subspace methods discussed in Chapters 2 and 3 and elsewhere can also be used. 95 Buy this book from SIAM at http://www.1.2.html.1.1) affects the convergence.1.1. The basic estimates We will discuss specific implementation issues in § 6. Theorem 6.1.. Direct analysis. Then there are δ and KI such that if xc ∈ B(δ). Chapter 6 Inexact Newton Methods Theorem 5. This electronic version is for personal use and may not be duplicated or distributed.4. See [145] and [175]. 6. and x+ = xc + s then (6. The proof and application of Theorem 6.1). This was the view taken in [55] where inexact Newton methods in which the step satisfies (6. s satisfies (6.com/SIAM/FR16. Another way to look at this is to ask how an approximate solution of the linear equation for the Newton step affects the iteration. We will present this result in two parts. Any approximate step is accepted provided that the relative residual of the linear equation is small. In § 6.1 should be compared to that of Theorem 5. We follow [69] and refer to the term ηc on the right hand side of (6.1 describes how errors in the derivative/function affect the progress in the Newton iteration. In most of this chapter we will focus on the approximate solution of the equation for the Newton step by GMRES. Such methods are not new. for example.1.2) e+ ≤ KI ( ec + ηc ) ec . 6. larger) choice of the parameter η.1. Let the standard assumptions hold. .1.1) as the forcing term.

Theorem 6.Copyright ©1995 by the Society for Industrial and Applied Mathematics. the proof is complete. . and • if ηn ≤ Kη F (xn ) with q-order 1 + p.1). If we set KI = K + 4κ(F (x∗ )). This electronic version is for personal use and may not be duplicated or distributed.com/SIAM/FR16.1 hold.1 e+ ≤ ec − F (xc )−1 F (xc ) + 4κ(F (x∗ ))ηc ec ≤ K ec 2 + 4κ(F (x∗ ))ηc ec . and (4.1 for iterative methods in the next result.1. η ].2) note that if2 r = −F (xc )s − F (xc ) is the linear residual then s + F (xc )−1 F (xc ) = −F (xc )−1 r and (6.3.1. Let the standard assumptions hold.1 and Theorem 5. Moreover • if ηn → 0 the convergence is q-superlinear. (4. We summarize the implications of Theorem 6.2) and will suffice to explain most computational observations.2). Let δ be small enough so that the conclusions of Lemma 4. which is a direct consequence of (6. In [55] r = F (xc )s + F (xc ).6) imply that s + F (xc )−1 F (xc ) ≤ F (xc )−1 ηc F (xc ) ≤ 4κ(F (x∗ ))ηc ec . 96 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Proof.html. using (6. Hence. Buy this book from SIAM at http://www.1. (6.3) e+ = ec + s = ec − F (xc )−1 F (xc ) − F (xc )−1 r. then the inexact Newton iteration ¯ ¯ xn+1 = xn + sn . Now.1. where F (xn )sn + F (xn ) ≤ ηn F (xn ) converges q-linearly to x∗ . p for some Kη > 0 the convergence is q-superlinear 2 Our definition of r is consistent with the idea that the residual in a linear equation is b − Ax.7). To prove the first assertion (6.3) and Theorem 5. Then there are δ and η such that if x0 ∈ B(δ).2. where K is the constant from (5. {ηn } ⊂ [0.ec-securehost.

Buy this book from SIAM at http://www.2 that if F (x∗ ) is ill conditioned very small forcing terms must be used. x+ = xc + s. 6. Reduce δ and η if needed so that ¯ KI (δ + η ) < 1.1. This electronic version is for personal use and may not be duplicated or distributed.2) holds for xc ∈ B(δ). Then there is δ such that if xc ∈ B(δ). . ¯ This proves q-linear convergence with a q-factor of KI (δ + η ).1. Let δ be small enough so that (6.5) Now.1. ¯ 1−p p + Kη 2p F (x∗ ) p ) en 1+p Proof. To prove (6. ¯ If ηn → 0 then q-superlinear convergence follows from the definition. and ηc ≤ η < η < 1.1). M0 = (6.2). then ¯ (6. Then if n ≥ 0 and xn ∈ B(δ) we have en+1 ≤ KI ( en + ηn ) en ≤ KI (δ + η ) en < en . This is not the case and the purpose of this subsection is to describe more accurately the necessary restrictions on the forcing terms. Since KI = O(κ(F (x∗ ))). 1) other than requiring that 1 not be an accumulation point.1 implies that F (xc ) ≤ F (x∗ )ec + Since we have. F (x∗ )ec ec = F (x∗ )−1 F (x∗ )ec ≤ F (x∗ )−1 γ F (x∗ )−1 2 F (xc ) ≤ (1 + M0 δ) F (x∗ )ec . one might conclude from Theorem 6.3.7) and (6. s satisfies (6. Weighted norm analysis. ¯ where KI is from (6. If ηn ≤ Kη F (xn ) then we may use (4. with γ ec 2 2 . The results here differ from those in § 6.2) to conclude en+1 ≤ KI ( en which completes the proof.Copyright ©1995 by the Society for Industrial and Applied Mathematics.2. INEXACT NEWTON METHODS 97 Proof. Theorem 6.4) note that Theorem 4.html. Let the standard assumptions hold.1 in that no restriction is put on the sequence {ηn } ⊂ [0.0.com/SIAM/FR16.4) F (x∗ )e+ ≤ η F (x∗ )ec .ec-securehost.1. F (x∗ )e+ = F (x∗ )(ec + s) = F (x∗ )(ec − F (xc )−1 F (xc ) − F (xc )−1 r).

98 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS By Theorem 5.1 F (x∗ )(ec − F (xc )−1 F (xc )) ≤ K F (x∗ ) ec 2 .6) Since F (x∗ )F (xc )−1 r ≤ r + (F (x∗ ) − F (xc ))F (xc )−1 r ≤ (1 + 2γ F (x∗ )−1 we may set ec ) r .2). The application of Theorem 6.3 asserts q-linear convergence in the weighted norm if the initial iterate is sufficiently near x∗ .5 for all n) that try to minimize the number of inner iterations are completely justified by this result if the initial iterate is sufficiently near the solution.1. Note that the distance δ from the initial iterate to the solution may have to be smaller for (6. M1 = 2γ F (x∗ )−1 .4) to hold than for (6. This electronic version is for personal use and may not be duplicated or distributed. and M2 = K F (x∗ ) and obtain.1. The importance of this theorem for implementation is that choices of the sequence of forcing terms {ηn } (such as ηn = .4. Hence.com/SIAM/FR16. We make such a choice in one of the examples considered in § 6.html.7) and (6.4) is remarkable in its assertion that any method that produces a step with a linear residual less than that of the zero step will reduce the norm of the error if the error is measured with the weighted norm · ∗ = F (x∗ ) · . using (4.1.1.ec-securehost.2 in that we do not demand that {ηn } be bounded away from 1 by a sufficiently large amount. F (x∗ )e+ (6.Copyright ©1995 by the Society for Industrial and Applied Mathematics.4 and compare it to a modification of a choice from Buy this book from SIAM at http://www. . However (6. just that 1 not be an accumulation point.5).4 differs from Theorem 6.1.3 described in Theorem 6. In fact. F (x∗ )e+ ≤ F (x∗ )F (xc )−1 r + K F (x∗ ) ec 2 . Now let δ be small enough so that ¯ (1 + M1 δ)(1 + M0 δ)η + M2 δ F (x∗ )−1 ≤ η and the proof is complete. This is made precise in Theorem 6.7) ≤ (1 + M1 δ) r + M2 δ ec ≤ (1 + M1 δ)(1 + M0 δ)ηc F (x∗ )ec +M2 δ F (x∗ )−1 F (x∗ )ec ≤ ((1 + M1 δ)(1 + M0 δ)η + M2 δ F (x∗ )−1 ) F (x∗ )ec . Theorem 6. (6.1.

1.1 and leave the proof to the reader in Exercise 6. Our proof does not rely on the (possibly false) assertions that en < δ for all n or that xn → x∗ q-linearly with respect to the unweighted norm. The proof uses both (6.4) implies that if en ∗ < δ∗ then (6.2.2. Theorem 6. We state this result as Proposition 6. To prove the assertions on q-superlinear convergence. e < δ∗ = F (x∗ ) δ0 ∗ Set δ = F (x∗ ) −1 δ∗ .4 shows that the sequence of forcing terms need only be kept bounded away from Buy this book from SIAM at http://www.5.2 might lead one to believe that a small value of ηn is necessary for convergence.ec-securehost. We close this section with some remarks on the two types of results in Theorems 6. Rather we note that (6. η] with η < η < 1. The superlinear convergence results then follow exactly as in the proof of Theorem 6. Then there is δ such that if x0 ∈ B(δ). While Theorem 6.2).4.2 and 6. and • if ηn ≤ Kη F (xn ) p for some Kη > 0 the convergence is q-superlinear with q-order 1 + p. Then e0 ∗ < δ∗ if e0 < δ.1. {ηn } ⊂ [0.4) holds for ec < δ0 .8) Note that e < δ0 if x ∗ ≤ F (x∗ ) x ≤ κ(F (x∗ )) x ∗ . Our first task is to relate the norms · and · ∗ so that we can estimate δ. (6. Proof.1.1.1. . where F (xn )sn + F (xn ) ≤ ηn F (xn ) converges q-linearly with respect to · ∗ to x∗ .Copyright ©1995 by the Society for Industrial and Applied Mathematics. INEXACT NEWTON METHODS 99 [69] that decreases ηn rapidly with a view toward minimizing the number of outer iterations.com/SIAM/FR16. For all x ∈ RN . q-linear convergence with respect to the weighted norm · ∗ is equivalent to q-linear convergence of the sequence of nonlinear residuals { F (xn ) }. This proves the first assertion.1. Moreover • if ηn → 0 the convergence is q-superlinear.1 hold. eventually en will be small enough so that the conclusions of Theorem 6. note that since xn → x∗ . This electronic version is for personal use and may not be duplicated or distributed. Theorem 6. Let δ0 be such that (6.4. Let the standard assumptions hold.1.html. The implications for the choice of the forcing terms {ηn } are also different.1. then the inexact Newton ¯ iteration xn+1 = xn + sn .4) and (6. The most significant difference is the norm in which q-linear convergence takes place.9) en+1 ∗ ≤ η en ¯ ∗ < δ∗ and hence xn → x∗ q-linearly with respect to the weighted norm.

11) e+ ≤ KI (( ec + ηc ) ec + (xc ) ). then ¯ there is KE such that (6. We leave the proofs. Newton-CG.1. Buy this book from SIAM at http://www.com/SIAM/FR16.4 we will scale the l2 norm by a factor of 1/N so that the results for the differential and integral equations will be independent of the computational mesh.html. we state two theorems on inexact local improvement. ¯ 6. Newton-SOR. s satisfies (6. for example. s satisfies (6. Proposition 6.3. but at a price of a greater number of outer iterations. The name of the method indicates which linear iterative method is used. In this section we use the l2 norm to measure nonlinear residuals. Note that errors in the derivative. In the examples discussed in § 6. or GMRES to solve the linear system.1.e. terminating when the relative linear residual is smaller than ηc (i.1.1 and 6. [136] in the context of optimization. These methods have also been called truncated Newton methods [56].6. then (6. can be regarded as part of the inexact solution of the equation for the Newton step. Let the standard assumptions hold. Newton-iterative methods A Newton-iterative method realizes (6.10).2.5 may well suffice for linear convergence. This electronic version is for personal use and may not be duplicated or distributed.2. Then there is δ such that if xc ∈ B(δ). Errors in the function. 6. Then there are δ and KI such that if xc ∈ B(δ). 100 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. when (6.12) F (x∗ )e+ ≤ η F (x∗ )ec + KE (xc ) . Let the standard assumptions hold. This naming convention is taken from [175] and [145]. such as those arising from a difference approximation of the action of F (xc ) on a vector.1.3. See [55] or Proposition 6. .3. Theorem 6. In this section.1.10) F (xc )s + F (xc ) + (xc ) ≤ ηc F (xc ) + (xc ) and x+ = xc + s.1 and its proof for an illustration of this point. would use SOR.5. A constant sequence such as ηn = . and ηc ≤ η < η < 1. Let the standard assumptions hold and let xn → x∗ .1. Newton-GMRES. Then F (xn ) converges q-linearly to 0 if and only if en ∗ does.1) holds). to the reader as exercises. CG.ec-securehost. We will discuss other factors in the choice of forcing terms in § 6. which are direct analogs to the proofs of Theorems 6.1) with an iterative method for the linear system for the Newton step. x+ = xc + s. Theorem 6. The reason for this is that the Krylov methods use the scalar product and the estimates in Chapters 2 and 3 are in terms of the Euclidean norm.1. Typically the nonlinear iteration that generates the sequence {xn } is called the outer iteration and the linear iteration that generates the approximations to the steps is called the inner iteration.

In fact. Algorithm 6. the action of F (xc ) on a vector w is approximated by a forward difference. a sequence of approximate steps {sk } is produced. k = 0 2. [24]. . We provide a detailed analysis of the NewtonGMRES iteration. h. INEXACT NEWTON METHODS 101 6. If the linear iterative method is any one of the Krylov subspace methods discussed in Chapters 2 and 3 then each inner iteration requires at least one evaluation of the action of F (xc ) on a vector. . ρ = r 2 . β = ρ.2. Note that matrixvector products are replaced by forward difference approximations to F (x). v1 = r/ r 2 . Newton GMRES.2. application of GMRES to the linear equation for the Newton step with matrix-vector products approximated by finite differences is the same as the application of GMRES to the matrix Gh F (x) whose last k − 1 columns are the vectors vk = Dh F (x : vk−1 ) The sequence {vk } is formed in the course of the forward-difference GMRES iteration.ec-securehost. k T i. It is important to note that this is entirely different from forming the finite difference Jacobian ∇h F (x) and applying that matrix to w. F.com/SIAM/FR16. x.1.1. (5. r = −F (x). s = 0. that b = −F (x). Buy this book from SIAM at http://www. . . hjk = vk+1 vj ii. s = Vk y k . . We give a MATLAB implementation of this algorithm in the collection of MATLAB codes. .Copyright ©1995 by the Society for Industrial and Applied Mathematics. 0.k = vk+1 2 2 2 and k < kmax do (d) vk+1 = vk+1 / vk+1 (1. While ρ > η F (x) (a) k = k + 1 (b) vk+1 = Dh F (x : vk ) for j = 1. η. In many implementations [20]. s = −F (x)−1 F (x). 0)T (e) e1 = ∈ Rk+1 Minimize βe1 − Hk y k Rk+1 to obtain y k ∈ Rk . . an evaluation of the action of F (xc )T on a vector is also required. We begin by discussing the effects of a forward difference approximation to the action of the Jacobian on a vector. Rk+1 .15). . as pointed out in [20]. This electronic version is for personal use and may not be duplicated or distributed. we give the forward-difference GMRES algorithm for computation of the Newton step. To illustrate this point. and that the initial iterate for the linear problem is the zero vector.html. In the case of CGNR and CGNE. (f) ρ = βe1 − Hk y k 3. In our MATLAB implementation we solve the least squares problem in step 2e by the Givens rotation approach used in Algorithm gmres. fdgmres(s. vk+1 = vk+1 − hjk vj (c) hk+1. kmax. ρ) 1. Dh F (x : w) for some h.

The action of Gh F on Kk . we have. .1. where {vj }k is the orthonormal basis for Kk generated j=1 by Algorithm fdgmres. Consider the linear map defined by its action on {uj } Buj = Dh F (x : uj ). and δ such that if x ∈ B(δ). which is a special case of the more general results in [20]. ¯ (6. h ≤ h. Let the standard assumptions hold. This electronic version is for personal use and may not be duplicated or distributed.15) Assume that F (x)s + F (x) 2 2 ≤ η F (x) 2 ≤ η F (x) 2 + h¯ s 2 /2. for the purposes of ⊥ analysis. h.14) B − F (x) 2 ≤ h x 2 γ/2 ≤ h¯ /2 γ Since the linear iteration (fdgmres) terminates in k < kmax iterations.3. First let δ be small enough so that B(δ) ⊂ Ω and the conclusions to Lemma 4.13) F (x)s + F (x) 2 < (η + CG h) F (x) 2 . and Algorithm fdgmres terminates with k < kmax then the computed step s satisfies (6. Proposition 6. γ ¯γ h¯ ≤ F (x∗ )−1 −1 2 /2 Buy this book from SIAM at http://www.2.com/SIAM/FR16. Note that Dh F (x : uj ) = 1 0 F (x + th x 2 uj )uj dt 1 0 (F = F (x)uj + (x + th x 2 uj ) − F (x))uj dt. the difference between ∇h F and Gh F is that the columns are computed by taking directional derivatives based on two different orthonormal bases: ∇h F using the basis of unit vectors in coordinate directions and Gh F the basis for the Krylov space Kk ⊥ constructed by algorithm fdgmres.html. 102 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS As should be clear from inspection of Algorithm gmres. Since F is Lipschitz continuous and {uj } is an orthonormal basis for RN we have. the orthogonal complement of Kk . Hence Gh F is also first order accurate in h. Assuming that there is no error in the evaluation of F we may summarize the observations above in the following proposition. Bs + F (x) and therefore (6. is not used by the algorithm and. could be specified by the action of ∇h F on any basis for Kk .Copyright ©1995 by the Society for Industrial and Applied Mathematics. Let η ∈ (0.ec-securehost. Proof. 1). since B and Gh F agree on Kk . ¯ ¯ Then there are CG .1 hold. with γ = γ( x∗ 2 + δ). Let {uj } be any orthonormal basis for RN such that uj = vj for 1 ≤ j ≤ k.

Therefore.2.ec-securehost. σ ] ¯ with 0 < σ < 1 then the forward difference Newton-GMRES iteration ¯ xn+1 = xn + sn Buy this book from SIAM at http://www. In an implementation. Theorem 6.1. . The proofs are immediate consequences of Theorems 6.2. Then there are δ.1.2.1 applies to a restarted GMRES because upon successful termination (6.2. Note also that Proposition 6.4. and Proposition 6.1 hold. Then there is δ such that if x0 ∈ B(δ) and the sequences {ηn } and {hn } satisfy σn = ηn + CG hn ∈ [0. • if σn → 0 the convergence is q-superlinear.2.15) imply F (x∗ )−1 −1 2 s 2 /2 ≤ F (x)−1 ≤ F (x)s 2 −1 2 s 2 2 ≤ (1 + η) F (x) 2 ¯γ + h¯ s 2 /2. kmax. (6.15) completes the proof with CG = 4¯ (1 + ∗ )−1 . xn .1 hold.1. We summarize our results so far in the analogs of Theorem 6.16) with (6.Copyright ©1995 by the Society for Industrial and Applied Mathematics.3. ηn .2. F.4 and Theorem 6. 6.17) F (xn )sn + F (xn ) 2 < σn F (xn ) 2 .com/SIAM/FR16.1.1 and (6. we must specify not only the sequence {ηn }. γ Combining (6.16) s 2 ≤ 4(1 + η) F (x∗ )−1 F (x) 2 .2.1.1 and are left as (easy) exercises. Moreover.2. therefore.15) will hold. but also the steps {hn } used to compute forward differences.html. hn .1 implies that a finite difference implementation will not affect the performance of Newton-GMRES if the steps in the forward difference approximation of the derivatives are sufficiently small. Assume that the assumptions of Proposition 6. and • if σn ≤ Kη F (xn ) p for some Kη > 0 the convergence is q-superlinear 2 with q-order 1 + p. INEXACT NEWTON METHODS 103 Then Lemma 4. This electronic version is for personal use and may not be duplicated or distributed. σ such that if x0 ∈ B(δ) and the sequences {ηn } and {hn } ¯ satisfy σn = ηn + CG hn ≤ σ ¯ then the forward difference Newton-GMRES iteration xn+1 = xn + sn where sn is computed by Algorithm fdgmres with arguments (sn . Theorem 6. η) F (x 2 Proposition 6.2. ρ) converges q-linearly and sn satisfies (6. Assume that the assumptions of Proposition 6.2.

2. may not converge rapidly enough to be useful. and • if σn ≤ Kη F (xn ) p for some Kη > 0 the convergence is q-superlinear 2 with q-order 1 + p. [111]. The choice in Buy this book from SIAM at http://www. If a constant η is too small. If a transpose is needed. Other iterative methods may be used as the solver for the linear equation for the Newton step.13. . xn . Other Newton-iterative methods. and [12]. [63]. as it is in the context of unconstrained minimization. the function.Copyright ©1995 by the Society for Industrial and Applied Mathematics. but the choice of that constant depends on the problem.4. [78].ec-securehost. See also [6]. The reader may explore this further in Exercise 6.2. 104 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS where sn is computed by Algorithm fdgmres with arguments (sn . a forward difference formulation is not an option because approximation of the transpose-vector product by differences is not possible. The reader may also be interested in experimenting with the other Krylov subspace methods described in § 3.com/SIAM/FR16.2 states that the observed behavior of the forward difference implementation will be the same as that of an implementation with exact derivatives. ρ) converges q-linearly with respect to · ∗ and sn satisfies (6. In the case where hn is approximately the square root of machine roundoff.6. σn ≈ ηn if ηn is not too small and Theorem 6. We adjust η as the iteration progresses with a variation of a choice from [69]. In addition.3. Newtonmultigrid [99] is one approach using a stationary iterative method.5. This issue is independent of the particular linear solver and our discussion is in the general inexact Newton method setting. GMRES may not be practical. As input we must give the initial iterate. but require evaluation of transpose-vector products.2. for a small m. hn .html. Newton-GMRES implementation We provide a MATLAB code nsolgm in the collection that implements a forward-difference Newton-GMRES algorithm. GMRES(m). CGNR and CGNE offer alternatives. Computation of ∇h F (x) and its transpose analytically is one option as is differentiation of F automatically or by hand. the vector of termination tolerances as we did for nsol. Moreover • if σn → 0 the convergence is q-superlinear. ηn . Newton-CG is a possible approach. When storage is restricted or the problem is very large. F. as it will be if CGNR or CGNE is used as the iterative method. This electronic version is for personal use and may not be duplicated or distributed. If F is symmetric and positive definite. much effort can be wasted in the initial stages of the iteration. Setting η to a constant for the entire iteration is often a reasonable strategy as we see in § 6. we must provide a method for forming the sequence of forcing terms {ηn }. Bi-CGSTAB and TFQMR should be considered in all cases where there is not enough storage for GMRES(m) to perform well. kmax. 6.17). This algorithm requires different inputs than nsol. 6.

Copyright ©1995 by the Society for Industrial and Applied Mathematics.9 as recommended in [69]. There is a chance that the final iterate will reduce F far beyond the desired level and that the cost of the solution of the linear equation for the last step will be more accurate than is really needed. (6. We use dirder to approximate directional derivatives and use the default value of h = 10−7 .1. The constant . B It may happen that ηn is small for one or more iterations while xn is still far from the solution.9 asks the reader to modify nsolgm so that other choices of the sequence {ηn } can be used.ec-securehost. 2 n > 0. This electronic version is for personal use and may not be duplicated or distributed. γηn−1 ≤ .1. Over solving means that the linear equation for the Newton step is solved to a precision far beyond what is needed to correct the nonlinear iteration.18) B ηn =   ηmax . γηn−1 > . In order to specify the choice at n = 0 and bound the sequence away from 1 we set (6. ηn ). The idea is that if ηn−1 is sufficiently large we do not let ηn decrease by much more than a factor of ηn−1 . 1] is a parameter. 2 A 2 min(ηmax . n > 0.            n = 0. Exercise 6.20) C ηn = min(ηmax . max(ηn . A for n > 0 would guarantee q-quadratic convergence by then setting ηn = ηn Theorem 6. γηn−1 )). the most information possible would be extracted from the inner iteration.9 and ηmax = . n > 0.9999 are used. INEXACT NEWTON METHODS 105 [69] is motivated by a desire to avoid such over solving. This oversolving on the final step can be controlled comparing the norm of the current nonlinear residual F (xn ) to the nonlinear residual norm at which the iteration would terminate τt = τa + τr F (x0 ) and bounding ηn from below by a constant multiple of τt / F (xn ) .com/SIAM/FR16.1.1 is somewhat arbitrary.   min(η A max . As a measure of the degree to which the nonlinear iteration approximates the solution we begin with A ηn = γ F (xn ) 2 / F (xn−1 ) 2 . .5τt / F (xn ) )).18) the parameter ηmax is an upper limit on the sequence {ηn }. This safeguarding does improve the performance of the iteration.1. . If ηn is uniformly bounded away from 1.5. max(ηn . In all the examples in this book we use the value γ = .  n = 0. Buy this book from SIAM at http://www.html. ηn ). In (6. In this way. The algorithm nsolgm does this and uses (6. A method of safeguarding was suggested in [69] to avoid volatile decreases in ηn . In [69] the choices γ = . A where γ ∈ (0.   ηmax .19) C ηn = A min(ηmax .

τ. Note that if only a single inner iteration is needed. nsolgm(x.4 are independent of the outer iteration. We revisit the H-equation and solve a preconditioned nonlinear convection-diffusion equation. the cost of the entire iteration can be greatly reduced. (b) fdgmres(s. Examples for Newton-GMRES In this section we consider two examples. We illustrate the consequences of this in § 6. and the Krylov basis {vk }m .5. x + hv. One new function evaluation will be needed to approximate the action of F (xc ) on a vector Buy this book from SIAM at http://www. we also terminate the outer iteration on small l2 norms of the nonlinear residuals.9. k=1 Since GMRES forms the inner iterates and makes termination decisions based on scalar products and l2 norms. From these plots one can compare not only the number of outer iterations. Counts of function evaluations corresponding to outer iterations are indicated by circles.4. η) √ 1.1. rc = r+ ≤ τr r0 + τa exit.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Since the Newton steps for M F (x) = 0 are the same as those for F (x) = 0. F (x + hv). F (x). In all the figures we plot the relative nonlinear residual F (xn ) 2 / F (x0 ) 2 against the number of function evaluations required by all inner and outer iterations to compute xn . F. σ = r+ /rc . F. the total cost of the outer iteration will be two function evaluations since F (xc ) will be known. but also the total cost. s. 6. . rc = r0 = F (x) 2 / N √ 2. Do while F (x) 2 / N > τr r0 + τa (a) Select η. If a maximum of m GMRES iterations is needed (or GMRES(m) is used as a linear solver) the storage requirements of Algorithm nsolgm are the m + 5 vectors x. For large and poorly conditioned problems.com/SIAM/FR16.ec-securehost. x. GMRES will have to be restarted. Hence if a large value of η can be used. we scale the l2 norm of the nonlinear √ residual by a factor of 1/ N so that constant functions will have norms that are independent of the computational mesh. This electronic version is for personal use and may not be duplicated or distributed. because of our applications to differential and integral equations. it is equivalent to precondition the nonlinear equation before calling nsolgm. The preconditioners we use in § 6.3.html. This enables us to directly compare the costs of different strategies for selection of η.4 and ask the reader to make the necessary modifications to nsolgm in Exercise 6. η) (c) x = x + s (d) Evaluate F (x) (f) If F (x) 2 √ (e) r+ = F (x) 2 / N . However. 106 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Algorithm 6. Note that the cost of an outer iterate is one evaluation of F to compute the value at the current iterate and other evaluations of F to compute the forward differences for each inner iterate.

4 outer iterations.9 and ηmax = .20) to γ = .1 we plot the progress of the iteration using η = .6 incurred a cost of 3.com/SIAM/FR16. c = . and roughly 275 thousand floating-point operations.1 with the solid line and using the sequence given by (6. 6. which therefore has a minimum cost of three function evaluations. Here the iteration with ηn = . This cost is entirely a result of the computation and factorization of a single Jacobian. The chord method with the different (but very similar for this problem) l∞ termination condition for the same problem.25. We set the parameters in (6. Chandrasekhar H-equation.1 performed slightly better and Buy this book from SIAM at http://www. reported in § 5. In Fig.4.ec-securehost.25.9 using two schemes for selection of the parameter η. We used τr = τa = 10−6 in this example. The initial iterate is the function identically one. which terminated after 10 function evaluations. In Fig.html.3 million floating-point operations. INEXACT NEWTON METHODS 107 and then F (x+ ) will be evaluated to test for termination. This electronic version is for personal use and may not be duplicated or distributed.9. This is a slightly higher overall cost than the other approach in which {ηn } is given by (6. For each example we compare a constant value of ηn with the choice given in (6. Newton-GMRES for the H-equation. We solve the H-equation on a 100point mesh with c = . This choice of ηmax was the one that did best overall in our experiments on this problem.1.20).1. 6.20) and used as the default in nsolgm. slower by a factor of over 1000. We set the parameters in (6. We see that the constant η iteration terminates after 12 function evaluations.20) with the dashed line. 3 outer iterations.2 the results change for the more ill-conditioned problem with c = .20) to γ = . 10 10 10 10 10 10 10 10 10 0 -1 -2 Relative Nonlinear Residual -3 -4 -5 -6 -7 -8 0 2 4 6 8 Function Evaluations 10 12 Fig. 6. 6. .9999. and 230 thousand floating-point operations.Copyright ©1995 by the Society for Industrial and Applied Mathematics.9 and ηmax = . We also count the evaluation of F (x0 ) as an additional cost in the evaluation of x1 .

and 537 thousand floating-point operations.9999 10 0 Relative Nonlinear Residual 10 -1 10 -2 10 -3 10 -4 0 2 4 6 8 10 12 Function Evaluations 14 16 18 Fig. This electronic version is for personal use and may not be duplicated or distributed.2.4. 6.2. tial equation (6. 108 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 0 10 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 10 -6 10 -7 0 5 10 15 Function Evaluations 20 25 Fig. Newton-GMRES for the H-equation. 6. and roughly 513 thousand floating-point operations.ec-securehost. c = . 6. The iteration with the decreasing {ηn } given by (6.21) We consider the partial differen- − ∇2 u + Cu(ux + uy ) = f Buy this book from SIAM at http://www. Newton-GMRES for the PDE. C = 20.html.20) required 23 function evaluations.3. 7 outer iterations. .Copyright ©1995 by the Society for Industrial and Applied Mathematics. terminated after 22 function evaluations and 7 outer iterations.com/SIAM/FR16. Convection-diffusion equation.

This electronic version is for personal use and may not be duplicated or distributed.9 the reader is asked to explore this. When preconditioning was not done. The iteration with {ηn } given by (6.ec-securehost.3.21) on a 31 × 31 grid using centered differences. the choice of {ηn } given by (6. To be consistent with the second-order accuracy of the difference scheme we used τr = τa = h2 . . a large choice of ηmax or constant η may be the more efficient choice.5.20) we set γ = . INEXACT NEWTON METHODS 109 with homogeneous Dirichlet boundary conditions on the unit square (0.21) with the fast Poisson solver fish2d. In Exercise 6. 4 outer iterations. the iteration for C = 20 required more than 80 function evaluations to converge. Preconditioning is crucial for good performance. 1). Buy this book from SIAM at http://www.com/SIAM/FR16. In cases where the initial iterate is very good and only a single Newton iterate might be needed. As in § 3. 1) × (0. In this case. however. We compare the same possibilities for {ηn } as in the previous example and report the results in Fig. Examples of such cases include (1) implicit time integration of nonlinear parabolic partial differential equations or ordinary differential equations in which the initial iterate is either the converged solution from the previous time step or the output of a predictor and (2) solution of problems in which the initial iterate is an interpolant of the solution from a coarser mesh.html.5 ).5.9 and ηmax = . and 2.6 million floating-point operations.20) required 16 function evaluations. We set C = 20 and u0 = 0. The computation with constant η terminated after 19 function evaluations and 4 outer iterates at a cost of roughly 3 million floating-point operations.7 we discretized (6.Copyright ©1995 by the Society for Industrial and Applied Mathematics. where h = 1/32. In the computation we preconditioned (6. f has been constructed so that the exact solution was the discretization of 10xy(1 − x)(1 − y) exp(x4.20) reduced the number of inner iterations in the early stages of the iteration and avoided oversolving. 6. In (6.

Exercises on inexact Newton methods 6.5.10.2. Prove Theorems 6. 10.2.1.5.75 affect the results? Present the results of your experiments as graphical iteration histories. How do the choices ηn = 1/(n + 1).4. 0 < < 1. 40 and how preconditioning affects the iteration. For the H-equation example in § 6.6.1.1. 6. What happens when c = 1? See [119] for an explanation of this. (n + 2)−1 ) [56]. Verify (6. what are some possible implications? 6.12. 6.2. ηn = 21−n [24].5. [23]. 6.5.8. and ηn = . how is the performance affected if GMRES(m) is used instead of GMRES? Buy this book from SIAM at http://www.5. In the DASPK code for integration of differential algebraic equations. Prove Theorems 6. and for the convectiondiffusion equation various values of C such as C = 1.5. 110 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 6.11. ηn = min( F (xn ) 2 . 6.ec-securehost. vary the parameter c and see how the performance of Newton-GMRES with the various choices of {ηn } is affected.14). For the H-equation experiment with various values of c. which may be less often that with each outer iteration. Can anything like Proposition 6.5.4 equally accurate? 6.7.9. such as c = .99999. 6.6.html.1 by showing that if the standard assumptions hold.4. Verify (6. ηn = 10−4 [32].1 be proved for a finite-difference BiCGSTAB or TFQMR linear solver? If not.4.8). and x is sufficiently near x∗ then F (x) /(1 + ) ≤ F (x∗ )e ≤ (1 + ) F (x) .3.5. Is there a relation to the Shamanskii method? 6.5. 6.05.5.com/SIAM/FR16. This electronic version is for personal use and may not be duplicated or distributed. .1. Give an example in which F (xn ) → 0 q-linearly but xn does not converge to x∗ q-linearly.4. Discuss this strategy and when it might be useful.5.5. preconditioned Newton-GMRES is implemented in such a way that the data needed for the preconditioner is only computed when needed. 6. Experiment with the sequence of forcing terms {ηn }. Are the converged solutions for the preconditioned and unpreconditioned convection-diffusion example in § 6. ηn = . 6.5. Prove Proposition 6. .5.999999. Duplicate the results in § 6.5 and 6.5. For the convection-diffusion example in § 6.Copyright ©1995 by the Society for Industrial and Applied Mathematics.2. .1.2 and 6.

Buy this book from SIAM at http://www. and nsola. If you have done Exercise 5. floatingpoint operations.25.html.Copyright ©1995 by the Society for Industrial and Applied Mathematics.7. The MATLAB codes fdkrylov. 6. which allows you to choose from a variety of forward difference Krylov methods.ec-securehost. Apply Newton-CGNR.5. apply nsolgm to the same two-point boundary value problem and compare the performance. How will you deal with the need for a transpose in the case of CGNR? How does the performance (in terms of time.com/SIAM/FR16.5. INEXACT NEWTON METHODS 111 6.13.4. . Newton-Bi-CGSTAB. function evaluations) compare to Newton-GMRES? Experiment with the forcing terms. which we describe more fully in Chapter 8. or Newton-TFQMR to the two examples in § 6.14. This electronic version is for personal use and may not be duplicated or distributed. might be of use in this exercise.

This electronic version is for personal use and may not be duplicated or distributed.html. .ec-securehost.Copyright ©1995 by the Society for Industrial and Applied Mathematics.com/SIAM/FR16. 112 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Buy this book from SIAM at http://www.

After the computation of x+ . In fact. In one dimension. is that solution of equations with the quasi-Newton approximation to the Jacobian is often much cheaper than using F (xn ) as the coefficient matrix. . is locally superlinearly convergent.Copyright ©1995 by the Society for Industrial and Applied Mathematics. then (7. This electronic version is for personal use and may not be duplicated or distributed. there is no cost in function evaluations associated with an inner iteration. In comparison with Newton-iterative methods. The advantages of such methods. For equations in several variables (7.3) alone is a system of N equations in N 2 113 Buy this book from SIAM at http://www. Broyden’s method [26] computes B+ by (7. quasi-Newton methods require only one function evaluation for each nonlinear iterate. The construction of B+ determines the quasi-Newton method. Broyden’s method. Bc is updated to form B+ . and hence is a very powerful alternative to Newton’s method or the chord method.3) B+ s = y. Broyden’s method is an example of a secant update. Chapter 7 Broyden’s method Quasi-Newton methods maintain approximations of the solution x∗ and the Jacobian at the solution F (x∗ ) as the iteration progresses.3.1) −1 x+ = xc − Bc F (xc ). Hence.3) uniquely specifies the classical secant method. The method we present in this chapter. the cost is little more than that of the chord method. as we shall see in § 7.2) y = F (x+ ) − F (xc ) and s = x+ − xc . sT s sT s In (7. these methods could have an advantage in terms of function evaluation cost over Newtoniterative methods. (7. This means that the updated approximation to F (x∗ ) satisfies the secant equation (7. If xc and Bc are the current approximate solution and Jacobian. if the Jacobian is dense.ec-securehost.com/SIAM/FR16.2) B+ = Bc + (y − Bc s)sT F (x+ )sT = Bc + .html. if a good preconditioner (initial approximation to F (x∗ )) can be found.

More subtle structural properties can also be preserved by secant methods. [95]. In this book we focus on Broyden’s method and our methods of analysis are not as general as those in [62] and [64]. We will express our results in terms of the error in the Jacobian (7. and [64] for discussion and analysis of several such methods. where B ≈ A is updated.com/SIAM/FR16. While we could use E = B −F (x). [62]. When indexing of the steps is necessary. One interesting result. The analysis is somewhat clearer in the linear case and the plan of this chapter is to present results for the linear case for both the theory and the examples. and [114] for discussions of quasi-Newton methods with inaccurate data. . [63].html. See [62] and [64] for discussion of this topic. which we will not discuss further. Ax = b. The Dennis–Mor´ condition e In this section we consider general iterative methods of the form (7. Some examples can be found in [101]. We must relate the superlinear convergence condition that ηn → 0 in Theorem 6. sn Buy this book from SIAM at http://www.4) is consistent with our use of e = x − x∗ .2 to a more subtle. [84].6) n→∞ lim En sn = 0. Locally superlinearly convergent secant methods can be designed to maintain the sparsity pattern. [60] on the sequence e of steps {sn } and errors in the Jacobian {En } (7. [143]. symmetry. [82]. [65]. [143]. Verification of q-superlinear convergence cannot be done by the direct approaches used in Chapters 5 and 6. [113]. In this chapter we will assume that the function values are accurate and refer the reader to [66].4) E = B − F (x∗ ) and the step s = x+ −xc . This electronic version is for personal use and may not be duplicated or distributed. we will write sn = xn+1 − xn . (7. In this chapter we show that if the data x0 and B0 are sufficiently good.1.5) −1 xn+1 = xn − Bn F (xn ) where Bn = F (x∗ )+En ≈ F (x∗ ) is generated by some method (not necessarily Broyden’s method).Copyright ©1995 by the Society for Industrial and Applied Mathematics. is that for linear problems Broyden’s method converges in 2N iterations [29]. and we refer the reader to [28]. and [116]. [104]. [130]. The role of the secant equation is very deep. 7. Broyden’s method is also applicable to linear equations [29]. [96]. of the approximate Jacobian. the Broyden iteration will converge q-superlinearly to the root.1.ec-securehost. [104]. [84]. This technique is based on verification of the Dennis–Mor´ condition [61]. or positive definiteness. [82]. but easier to verify condition in order to analyze Broyden’s method. 114 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS unknowns. One can also apply the method to linear least squares problems [84].

Then xn → x∗ q-superlinearly if and only if xn → x∗ and the Dennis–Mor´ condition (7. if xn → x∗ q-superlinearly. let {Bn } be a sequence of nonsingular N × N matrices.1. This electronic version is for personal use and may not be duplicated or distributed. Now. e Proof. and hence (7.1. by (7. Since −F (xn ) = Bn sn = F (x∗ )sn + En sn we have (7.6) holds. .8) implies that En sn ≤ (2 F (x∗ ) νn + γ en ) sn which implies the Dennis-Mor´ condition (7. Our formulation of the Dennis–Mor´ result is e a bit less general than that in [60] or [63].ec-securehost. The assumption that xn = x∗ for any n implies that the sequence (7.6). Let the standard assumptions hold. let x0 ∈ RN be given and let {xn }∞ be given n=1 by (7. are used to prove superlinear convergence for Broyden’s method in this chapter and many other quasi-Newton methods. We use the fundamental theorem of calculus and the standard assumptions to obtain F (x∗ )en − F (xn ) = and hence Therefore. then for n sufficiently large (7. See [61]. sn Buy this book from SIAM at http://www. e Conversely. BROYDEN’S METHOD 115 The main result in this section.8) En sn ≤ F (x∗ )en+1 + γ en 2 /2. Let µn = En sn . 1 0 (F (x∗ ) − F (x∗ + ten ))en dt F (x∗ )en − F (xn ) ≤ γ en 2 /2.7) (7.6) holds. [28].1 and its corollary for linear problems. assume that xn → x∗ and that (7.9) we have en+1 = νn en ≤ 2νn sn . Theorem 7. By (7. Theorem 7. and [64] for several more examples.Copyright ©1995 by the Society for Industrial and Applied Mathematics.html. [62].10) νn = en+1 / en is defined and and νn → 0 by superlinear convergence of {xn }.1.7) En sn = −F (x∗ )sn − F (xn ) = −F (x∗ )en+1 + F (x∗ )en − F (xn ).9) sn /2 ≤ en ≤ 2 sn .com/SIAM/FR16.5). Assume that xn = x∗ for any n.

17) then (7. n=0 Let { n }∞ ⊂ RN be such that n=0 n n 2 Since xn → x∗ by assumption.13) holds. and we employ notation from all four of these references. (7. . and [115].11) (7. 116 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS We have.2.ec-securehost.11) again. [104].html.14) Moreover.18) T ψn+1 = ψn − θn (ηn ψn )ηn + n is either 1 or 0 for lim η T ψn n→∞ n = 0. Since ηn → 0. Buy this book from SIAM at http://www. If {ψn }∞ is given by n=1 (7. That paper was the basis for [168]. 2 and let {ηn }∞ be a set of vectors in RN such that ηn n=0 all n. This electronic version is for personal use and may not be duplicated or distributed. Combining (7. ηn → 0 where < ∞. 2 − θ).2. Convergence analysis Our convergence analysis is based on an approach to infinite-dimensional problems from [97].12) sn En sn = (Bn − F (x∗ ))sn = −F (xn ) − F (x∗ )sn . Hence.15) implies that (7.1. ˆ Lemma 7. using the standard assumptions and (7. F (xn ) + F (xn )sn (7. ηn = 2 F (x∗ )−1 (µn + γ en ).14) and (7. which is the inexact Newton condition.1. Since µn → 0 by assumption. Let ψ0 ∈ RN be given.Copyright ©1995 by the Society for Industrial and Applied Mathematics.com/SIAM/FR16.13) µn ≤ F (x∗ )−1 −1 /2.2. (7. sn ≤ 2 F (x∗ )−1 F (xn ) . for n sufficiently large. We assume now that n is large enough so that (7. Let 0 < θ < 1 and let ˆ ˆ {θn }∞ ⊂ (θ. since (7.15) ≤ F (xn ) + F (x∗ )sn + (F (x∗ ) − F (xn ))sn ≤ (µn + γ en ) sn . We begin with a lemma from [104]. 7. We will use the l2 norm throughout our discussion. F (x∗ )−1 −1 ≤ F (x∗ )sn ≤ En sn + F (xn ) = µn sn + F (xn ) . the proof is complete by Theorem 6.16) F (xn ) + F (xn )sn ≤ ηn F (xn ) .

This inequality is used often in the analysis of quasi-Newton methods [63]. 2 ψn 2 ≤ ψn Hence if ψn = 0 (7. BROYDEN’S METHOD 117 Proof. M n=0 T (ηn ψn )2 ≤ ψ0 2 2 − ψM +1 ˆ θ2 2 2 ≤ ψ0 2 . To prove the result for n = 0 we use the inequality (7. 2a which is valid for a > 0 and |b| ≤ a. 2 ψn 2 ( ψn θn (2 − θn ) 2 − ψn+1 2 + n 2 ). From (7.21) Hence (7. This electronic version is for personal use and may not be duplicated or distributed.19) ∞ n=0 T (ηn ψn )2 < ∞.com/SIAM/FR16.21) we conclude that ψn+1 2 ≤ µ. for any M > 0. From (7. ≤ ψn ≤ ψn 2 2 2 2 T − θn (2 − θn )(ηn ψn )2 ˆ T − θ2 (ηn ψn )2 ≤ ψn 2 . 2 Therefore.ec-securehost.20) a2 − b2 ≤ a − b2 .18).22) T (ηn ψn )2 ≤ − ψn+1 2 ≤ ψn 2 − T θn (2 − θn )(ηn ψn )2 + 2 ψn 2 n 2. ˆ θ2 We let M → ∞ to obtain (7.html.19) implies convergence to zero of the terms in the series. which is (7. We first consider the case in which n = 0 for all n.Copyright ©1995 by the Society for Industrial and Applied Mathematics. in fact. Buy this book from SIAM at http://www. Convergence of the series in (7. which holds even if ψn = 0. . we can use the fact that ˆ θn (2 − θn ) > θ2 > 0 to show that the sequence {ψn } is bounded in l2 -norm by ψ0 ψn+1 2 2 2 and.20) we conclude that if ψn = 0 then T ψn − θn (ηn ψn )ηn 2 ≤ ψn 2 2 2 T − θn (2 − θn )(ηn ψn )2 T θn (2 − θn )(ηn ψn )2 . In that case.

≤ 2µ ˆ θ2 2µ ˆ θ2 M n=0 ( ψn 2 − ψn+1 2 + M n 2) = ψ0 2 − ψM +1 2+ n 2 n=0 ≤ 2µ2 . The standard assumptions in this case reduce to nonsingularity of A. Our first task is to show that the errors in the Broyden approximations to A do not grow.com/SIAM/FR16. 118 where ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS ∞ i 2 i=0 µ= Hence M T 2 n=0 (ηn ψn ) + ψ0 2 .2. The reader should compare the statement of the result with that in the nonlinear case. called bounded deterioration [28]. This electronic version is for personal use and may not be duplicated or distributed.2.23) B+ = Bc + θc (y − Bc s)sT (Ax+ − b)sT = Bc + θc . ˆ θ2 Hence (7. We may use Lemma 7. the Broyden update has a very simple form. .24) B+ = Bc + θc (y − Bc s)sT (Ec s)sT = Bc − θc .1 to prove a result from [130] on the convergence of Broyden’s method for linear problems. Then (7.2.2.23).19) holds and the proof is complete. s 2 s 2 2 2 Lemma 7. sT s sT s If xc and Bc are the current approximations to x∗ = A−1 b and A and s = x+ − xc then y − Bc s = (Ax+ − Axc ) − Bc (x+ − xc ) = −Ec s and therefore (7. In this section F (x) = Ax−b and Bn = A+En .1 is nicely illustrated by this application of the lemma. Buy this book from SIAM at http://www.25) E+ 2 ≤ Ec 2 . and xc ∈ RN be given. In the linear case.Copyright ©1995 by the Society for Industrial and Applied Mathematics. The role of ˆ the parameter θ and the sequence {θn } in Lemma 7. 2].2. Linear problems.html. We will consider a modified update (7.ec-securehost. 7. θc ∈ [0. Let A be nonsingular.1. This property. Let Bc be nonsingular and let B+ be formed by the Broyden update (7. is an important part of the convergence analysis.

θc can always be selected to make B+ nonsingular. 1 − sign(γc )σ θc =  . s 2 2 ≤ Ec 2 I − θ c Ps 2 and I − θc Ps 2 ≤ 1 by orthogonality of Ps and the fact that 0 ≤ θc ≤ 2. Specifically Proposition 7. Theorem 7.1 for some θ ∈ (0. e We leave the proof as Exercise 7. By subtracting A from both sides of (7. .ec-securehost. 1) is fixed.2.24) we have E+ = E c − θ c (Ec s)sT s 2 2 (7.2. This electronic version is for personal use and may not be duplicated or distributed.  1. However the results in [130] assume only that the ˆ sequence {θn } satisfies the hypotheses of Lemma 7.1. If {θn } is such that the matrices Bn obtained by (7.28) −1 xn+1 = xn − Bn (Axn − b) then the Dennis–Mor´ condition implies convergence of {xn } to x∗ = A−1 b.26) = Ec (I − θc Ps ).1 for some θ ∈ (0. 1 − γc where γc = −1 −1 (Bc y)T s (Bc As)T s = 2 s 2 s 2 2 and σ ∈ (0.23) are nonsingular and {xn } is given by the modified Broyden iteration (7. For linear problems one can show that the Dennis–Mor´ condition implies e convergence and hence the assumption that convergence takes place is not needed. Buy this book from SIAM at http://www.23) are nonsingular then the modified Broyden iteration (7. The superlinear convergence result is remarkable in that the initial iterate need not be near the root. Assume that {θn } satisfies ˆ the hypotheses of Lemma 7.html.28) converges q-superlinearly to x∗ = A−1 b for every initial iterate x0 .1.2. Let A be nonsingular. Now.1.2. 1). In [130] one suggestion was  |γc | ≥ σ. Let A be nonsingular. BROYDEN’S METHOD 119 Proof.1 for some θ ∈ (0.com/SIAM/FR16. where Ps is the orthogonal projector (7. 1).Copyright ©1995 by the Society for Industrial and Applied Mathematics. Assume that {θn } satisfies the ˆ hypotheses of Lemma 7.5. If {θn } is such the matrices Bn obtained by (7. otherwise. 1) and that θc is always chosen so that B+ is nonsingular.2.27) This completes the proof as E+ 2 Ps = ssT .

that the Broyden iteration will terminate in 2N steps in the linear case.1. Let φ ∈ RN be given.29) holds. 120 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Proof. sn 2 sn 2 This completes the proof. We say that the Broyden sequence ({xn }. that does not use the Frobenius norm and extends to various infinite-dimensional settings.30) T ηn ψ n = T En sn (En φ)T sn = φT → 0. x0 .2. ψn = En φ.1 with n =0 to conclude that (7. B0 ) exists if F is defined at xn for all n.1. This will imply q-superlinear convergence e by Theorem 7. Bn is nonsingular for all n. sn 2 2 and.com/SIAM/FR16.1. {Bn }) for the data (F. [63]. Nonlinear problems. Our analysis of the nonlinear case is different from the classical work [28]. and we may invoke Lemma 7.2. As indicated in the preface.26) T T En+1 φ = (I − θn Pn )En φ.html.1) and (7.2. However we must show that the sequence {Bn } of Broyden updates exists and remains nonsingular. Assuming that (7.29) n→∞ lim φT En sn sn 2 = 0. T ηn = sn / sn 2 . Buy this book from SIAM at http://www. [84]. shows that the j component of has limit zero.2. Our approach is to show that for every φ ∈ RN that (7. we have n→∞ lim En sn =0 sn 2 which is the Dennis–Mor´ condition. The result here is not a local convergence result.ec-securehost. .Copyright ©1995 by the Society for Industrial and Applied Mathematics. [115]. we provide a proof based on ideas from [97]. This electronic version is for personal use and may not be duplicated or distributed. [82]. the unit vector in the jth En sn coordinate direction. Let Pn = Since for all v ∈ RN and all n T v T (En φ) = φT (En v) sn sT n . [143]. and [104]. Definition 7. For nonlinear problems our analysis is local and we set θn = 1. and xn+1 and Bn+1 can be computed from xn and Bn using (7.2). It is global in the sense that q-superlinear convergence takes place independently of the initial choices of x and B.1 and Proposition 7. 7. by (7. It is known [29]. We make a definition that allows us to compactly express this fact. setting φ = ej .2. sn 2 As j was arbitrary.

this growth can be limited and therefore the Broyden sequence exists and the Broyden iterates converges to x∗ at least q-linearly. which. Unlike the linear case. together with the q-linear convergence e proved earlier completes the proof of local q-superlinear convergence. Buy this book from SIAM at http://www. the possible increase can be bounded in terms of the errors in the current and new iterates. Our first task is to show that bounded deterioration holds. Then (7. However in the nonlinear case. The development of the proof is complicated. We then show that if the initial data is sufficiently good. . Let the standard assumptions hold. (F (xc + ts) − F (x∗ ))s dt (∆c s)sT s 2 2 (F (xc + ts) − F (x∗ )) dt. B0 ) and that the Broyden iterates converge qsuperlinearly to x∗ . We first prove a bounded deterioration result like Lemma 7.2. Bounded deterioration.7) implies that Ec s = −F (x+ ) + (F (x+ ) − F (xc ) − F (x∗ )s) (7.com/SIAM/FR16. This electronic version is for personal use and may not be duplicated or distributed.32) as F (x+ ) = −Ec s + and then using (7. Theorem 7.31) E+ 2 ≤ Ec + γ( ec 2 + e+ 2 )/2 Proof.2.2.ec-securehost. the sequence { En } need not be monotone non-increasing. BROYDEN’S METHOD 121 We will show that if the standard assumptions hold and x0 and B0 are sufficiently good approximations to x∗ and F (x∗ ) then the Broyden sequence exists for the data (F. Note that (7.html.32) = −F (x+ ) + We begin by writing (7.2) to obtain E+ = Ec (I − Ps ) + where Ps is is given by (7. the errors in the Jacobian approximation can grow as the iteration progresses. Assume that −1 x+ = xc − Bc F (xc ) = xc + s ∈ Ω and B+ is given by (7.2).6).Copyright ©1995 by the Society for Industrial and Applied Mathematics.2. x0 . Finally we verify the Dennis–Mor´ condition (7. Let xc ∈ Ω and a nonsingular matrix Bc be given.27) and ∆c = 1 0 1 0 1 0 (F (xc + ts) − F (x∗ ))s dt. However.

2 for the linear case E+ and the proof is complete. By Theorem 7. + γ(1 + r)rn δ/2. To do this reduce δ if needed so that γ(1 + r)δ < δ1 (7. in the same way to obtain En+1 2 ≤ E0 2+ rj ≤ δB + which completes the proof. en+1 2 2 and therefore 2 En+1 ≤ En 2 + γ(1 + r) en 2 /2 ≤ En (1 + r)γδ 2 We can estimate En 2 . the q-factor is at most KA (δ + δ1 ) ≤ r. 2(1 − r) Buy this book from SIAM at http://www. Theorem 7. Since E0 2 < δB < δ1 we may begin the induction.3. n j=0 < δ1 .3. .com/SIAM/FR16.3 hold.2. by the standard assumptions ∆c 2 ≤ γ( ec ≤ Ec 2 + e+ 2 )/2. Then there are δ and δB such that if x0 ∈ B(δ) and E0 2 < δB the Broyden sequence for the data (F. .4.html. Now assume that Ek 2 < δ1 for all 0 ≤ k ≤ n. Then. Proof. 122 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Hence. Theorem 7.4.2. The proof is more complex than similar results in Chapters 5 and 6. B0 ) exists and xn → x∗ q-linearly with q-factor at most r. 1) be given. Local q-linear convergence.2. Let δ and δ1 be small enough so that the conclusions of Theorem 5.Copyright ©1995 by the Society for Industrial and Applied Mathematics.3 states that Broyden’s method is locally q-linearly convergent. 2 2 2 We show that En 2 ≤ δ1 by induction. Let the standard assumptions hold. .34) δB = δ1 − δ 2 . This electronic version is for personal use and may not be duplicated or distributed. where KA is from the statement of Theorem 5.ec-securehost. (1 + r)γδ ≤ δ1 . Let r ∈ (0.33) δ2 = 2(1 − r) and set (7. En+1 Since En 2 2 2 ≤ En ≤ r en 2 + γ( en+1 2 + en 2 )/2. + ∆c Just as in the proof of Lemma 7. En−1 2 .2. . which will prove the result. We must explore the relation between the size of E0 needed to enforce q-linear convergence and the q-factor. We will prove the result by choosing δB and reducing δ if necessary so that E0 2 ≤ δB will imply that En 2 ≤ δ1 for all n.2. x0 . reducing δ and δ1 further if needed.

Another approach. after the storage available for the iteration is exhausted. as in the proof of Theorem 7. BROYDEN’S METHOD 123 Verification of the Dennis–Mor´ condition. . x0 .2.2.com/SIAM/FR16. Then there are δ and δB such that if x0 ∈ B(δ) and E0 2 < δB the Broyden sequence for the data (F.1 implies that T ηn ψn = T En sn (En φ)T sn = φT → 0.html.3 and the standard assumptions imply ∆n 2 ≤ γ(1 + r)rn δ/2.1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. n We wish to apply Lemma 7.1 with T ψn = En φ. Let the standard assumptions hold.2. Proof. Implementation of Broyden’s method The implementation of Broyden’s method that we advocate here is related to one that has been used in computational fluid mechanics [73]. By Theorem 7.1. the oldest of the stored vectors is replaced by the most recent.2. and n = Pn ∆T φ. sn 2 2 (F (xn + tsn ) − F (x∗ )) dt. sn 2 sn 2 This. This electronic version is for personal use and may not be duplicated or distributed. Buy this book from SIAM at http://www.1 we need only show that Dennis–Mor´ condition (7. Let δ and δB be such that the conclusions of Theorem 7.1 let Pn = Set ∆n = 1 0 sn sT n . clears the storage and starts over. The final task is to verify e the Dennis–Mor´ condition. ηn = sn / sn 2 .ec-securehost. [142]. n The hypothesis in the lemma that n n 2 <∞ holds since Theorem 7. B0 ) exists and xn → x∗ q-superlinearly. Hence Lemma 7. 7.2. Note that T T En+1 φ = (I − Pn )En φ + Pn ∆T φ. restarting. e Theorem 7.6) holds e to complete the proof. Let φ ∈ RN be arbitrary. implies the Dennis–Mor´ condition and e completes the proof. Some such methods are called limited memory formulations in the optimization literature [138].3 hold. [31].3.2. As in the proof of Theorem 7.2.4. In these methods.

We begin by showing how the initial approximation to F (x∗ ). Equation (7. Then B + uv T is invertible if and only if 1 + v T B −1 u = 0. (B + uv T )−1 = I − 1 + v T B −1 u The expression (7. I) and xn = yn . v ∈ RN . This electronic version is for personal use and may not be duplicated or distributed.5. may be regarded as a preconditioner and incorporated into the defintion of F just as preconditioners were in the implementation of Newton-GMRES in § 6.36) −1 = xn − (B0 Cn )−1 F (xn ) = xn − Bn F (xn ) = xn+1 . x0 . x0 .35) holds for a given n.3. Let B be a nonsingular N × N matrix and let u. [177].3.com/SIAM/FR16. Proof. [11].35) for all n.2. We leave the proof to Exercise 7. Our plan will be to store Bn in a compact and easy to apply form. If (7. Buy this book from SIAM at http://www. Proposition 7. {Bn }) for the data (F.36). The consequence of this proof for implementation is that we can assume that B0 = I. In this case. Bn = B0 Cn (7. {Cn }) exists for the data −1 (B0 F. we can incorporate that into the −1 function evaluation. B0 ) exists. 124 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS This latter approach is similar to the restarted GMRES iteration. though a limited memory approach could be used as well.37) B −1 . We use a restarted approach. If a better choice for B0 exists. Then the Broyden sequence ({yn }. We proceed by induction.1.37) is called the Sherman–Morrison formula.1. Hence Bn+1 = Bn + F (xn+1 )sT n sn 2 2 −1 B0 F (yn+1 )sT n = B0 Cn+1 .html. (B −1 u)v T (7. Now sn = xn+1 − xn = yn+1 − yn by (7. . then −1 −1 yn+1 = yn − Cn B0 F (yn ) (7. Our basic implementation is a nonlinear version of the implementation for linear problems in [67]. [176].ec-securehost.4.35) holds by definition when n = 0. Lemma 7. sn 2 2 = B0 Cn + B0 This completes the proof. T Bn+1 = Bn + un vn . Assume that the Broyden sequence ({xn }. In the context of a sequence of Broyden updates {Bn } we have for n ≥ 0. B0 . The basis for this is the following simple formula [68].Copyright ©1995 by the Society for Industrial and Applied Mathematics.

if B0 = I.html. sn = −Bn F (xn ) = − j=0 Since the product n−2 j=0 T (I − wj vj )F (xn ) must also be computed as part of the computation of wn−1 we can combine the computation of wn−1 and sn as follows: w (7.41) 2 = − sn−1 2 wn−1 .Copyright ©1995 by the Society for Industrial and Applied Mathematics. vj }n−1 at a cost of O(N n) j=0 floating-point operations.ec-securehost. . Since the empty matrix product is the identity. . for n ≥ 0.39) (I − wj vj )F (xn ).38) is valid for n ≥ 0.40) = n−2 j=0 T (I − wj vj )F (xn ) 2 T + vn−1 w)−1 wn−1 = Cw w where Cw = ( sn−1 sn T = −(I − wn−1 vn−1 )w. the computation of the Broyden step) can be computed from the 2n vectors {wj . This electronic version is for personal use and may not be duplicated or distributed. BROYDEN’S METHOD 125 where un = F (xn+1 )/ sn Setting we see that. One may carry this one important step further [67] and eliminate the need to store the sequence {wn }. Note that (7. Therefore. one need only store the steps {sn } and their norms to construct the sequence {wn }. Buy this book from SIAM at http://www. wn = −sn+1 / sn 2 .38) = n−1 j=0 T (I − wj vj ).e. −1 = (I − w T T T Bn n−1 vn−1 )(I − wn−2 vn−2 ) .com/SIAM/FR16.. Moreover. (7. .42) −1 Bn = n−1 j=0 I+ sj+1 sT j sj 2 2 . (7. −1 Hence the action of Bn on F (xn ) (i.40) implies that for n ≥ 1 T −1 sn = −w + Cw w(vn−1 w) = w(−1 + Cw (Cw − sn−1 2 )) = −Cw w sn−1 Hence. we can write (7. In fact. −1 T −1 wn = (Bn un )/(1 + vn Bn un ) (7. (I − w0 v0 ) 2 and vn = sn / sn 2 . the Broyden step for the following iteration is n−1 −1 T (7.38) as (7.

τ.Copyright ©1995 by the Society for Industrial and Applied Mathematics. restarting with B0 when storage is exhausted. F. Algorithm 7. n − 1 z = z + sj+1 sT z/ sj j 2 2 Buy this book from SIAM at http://www. we solve (7. Our notation and algorithmic framework are based on [67].3. 126 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS We cannot use (7. making Broyden’s method competitive in storage with GMRES as a linear solver [67]. maxit.39) directly to compute sn+1 because sn+1 appears on both sides of the equation sn+1 = − I + (7. We use a restarting approach in the algorithm. n = −1.ec-securehost.42) and (7. for j = 0.3. 2. s0 = −F (x). r0 = F (x) 2 . z = −F (x) ii. Note that B0 = I is implicit in the algorithm.com/SIAM/FR16.42) and (7.1. itc = 0.1. this is consistent with setting the empty matrix product to the identity.44) to describe an algorithm. Our implementation is similar to GMRES(m).html. nmax) 1.44) sn+1 = − −1 Bn F (xn+1 ) −1 1 + sT Bn F (xn+1 )/ sn n 2 2 .6. We can use (7.43) for sn+1 to obtain (7. The storage requirement. Instead.44) −1 1 + sT Bn F (xn+1 )/ sn n 2 2 T −1 = 1 + vn Bn un is nonzero unless Bn+1 is singular. the denominator in (7.43) =− I+ sn+1 sT n sn 2 2 sn+1 sT n sn 2 2 n−1 j=0 I+ sj+1 sT j 2 sj 2 F (xn+1 ) −1 Bn F (xn+1 ). is of the same order as GMRES. of n + O(1) vectors for the nth iterate. This electronic version is for personal use and may not be duplicated or distributed. The termination criteria are the same as for Algorithm nsol in § 5. brsol(x. Our looping convention is that the loop in step 2(e)ii is skipped if n = 0. (e) if n < nmax then i. adding to the input list a limit nmax on the number of Broyden iterations taken before a restart and a limit maxit on the number of nonlinear iterations. Do while itc < maxit (a) n = n + 1. . By Proposition 7. itc = itc + 1 (b) x = x + sn (c) Evaluate F (x) (d) If F (x) 2 ≤ τr r0 + τa exit.

where z and F (x) may occupy the same j=0 storage location. the iteration will restart with xnmax .33) from § 3. s = −F (x). right-hand side. Examples for Broyden’s method Buy this book from SIAM at http://www.Copyright ©1995 by the Society for Industrial and Applied Mathematics. x. For linear problems. In Fig. and in several papers on infinite-dimensional problems as well [91]. we allow for increases in the nonlinear residual. We use the same mesh. [104].ec-securehost. However the dense matrix approach makes it possible to detect ill conditioning in the approximate Jacobians. Broyden’s method terminated after 9 iterations at a cost of roughly 1. .4.33) preconditioned with the Poisson solver fish2d. bounded deterioration implies that the Broyden matrices will be well conditioned and superlinear convergence implies that only a few iterates will be needed. where more elaborate numerical experiments that those given here are presented.com/SIAM/FR16. Our implementation of brsol in the collection of MATLAB codes stores z and F separately in the interest of clarity and therefore requires nmax + 3 vectors of storage. We use the MATLAB code brsol. coefficients.7. If the data is sufficiently good. The restarted iteration required 24 iterations and 3. √ In all the examples in this section we use the l2 norm multiplied by 1/ N . than the dense matrix approach proposed in [85]. We compare Broyden’s method without restarts (solid line) to Broyden’s method with restarts every three iterates (dashed line). in terms of both time and storage. One can also see highly irregular performance from the 7. 7.4. and solution as the example in § 3. If n < nmax. We set τa = τr = h2 where h = 1/32.5 million floating point operations. and therefore need storage for nmax + 2 vectors. As an example we consider the linear PDE (3. Linear problems. especially for large problems. [120]. This point has been explored in some detail in [67]. This requirement of one more vector than the algorithm in [67] needed is a result of the nonlinearity. If n = nmax. 7.7. The theory in this chapter indicates that Broyden’s method can be viewed as an alternative to GMRES as an iterative method for nonsymmetric linear equations. In view of all this. and F (x).html. requires O(nN ) floating-point operations and storage of n + 3 vectors. z.5 million floating point operations. A MATLAB implementation of brsol is provided in the collection of MATLAB codes. This electronic version is for personal use and may not be duplicated or distributed.1 we present results for (3. the steps {sj }n . a slightly higher cost than the GMRES solution. which computes xn and sn+1 . then the nth iteration of brsol. we feel that the approach of Algorithm brsol is the best approach.1. The Sherman–Morrison approach in Algorithm brsol is more efficient. sn+1 = z/(1 − sT z/ sn 2 ) n 2 (f) if n = nmax then n = −1. BROYDEN’S METHOD 127 iii. not compute snmax+1 .

Buy this book from SIAM at http://www. Broyden’s method for the linear PDE.ec-securehost.2) and c = . In the linear case [67] one can store only the residuals and recover the terminal iterate upon exit.4 illustrate the differences in the two methods. Our general recommendation is to try both. This electronic version is for personal use and may not be duplicated or distributed. Broyden’s method and Newton-GMRES both require more storage as iterations progress. Broyden’s method could be at a disadvantage as it might need to be restarted many times.9 (Fig. 7.com/SIAM/FR16. Hence. As before we solve the H-equation on a 100point mesh with c = . Chandrasekhar H-equation. 7.4. The cost in function evaluations for Broyden’s method is one for each nonlinear iteration and for Newton-GMRES one for each outer iteration and one for each inner iteration. The implementation without restarts required 153 thousand floating-point operations and the restarted method 150. a substantial improvement over the Newton-GMRES implementation. . We compare Broyden’s method without restarts (solid line) to Broyden’s method with restarts every three iterates (dashed line). The examples in this section.3). NewtonGMRES as the inner iteration progresses and Broyden’s method as the outer iteration progresses. For c = . and the sequence of steps saves a vector over the nonlinear case. if function evaluations are very expensive.html. Storage of the residual. 7. If storage is at a premium. 7. however. We consider the H-equation from Chapters 5 and 6 and a nonlinear elliptic partial differential equation. as well as those in § 8.2. the cost of a solution by Broyden’s method could be much less than that of one by Newton-GMRES. the vector z. and the number of nonlinear iterations is large. 10 1 10 Relative Residual 0 10 -1 10 -2 10 -3 0 5 10 Iterations 15 20 25 Fig. Nonlinear problems.9999 (Fig.Copyright ©1995 by the Society for Industrial and Applied Mathematics.1. 128 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS restarted method.9 both methods required six iterations for convergence.

10 10 10 10 10 10 10 10 10 0 -1 -2 Relative Nonlinear Residual -3 -4 -5 -6 -7 -8 0 1 2 3 Iterations 4 5 6 Fig.com/SIAM/FR16.Copyright ©1995 by the Society for Industrial and Applied Mathematics. The unrestarted iteration converges in 10 iterations at a cost of roughly 249 thousand floating-point operations. The restarted iteration requires 18 iterations and 407 floating-point operations. Broyden’s method for the H-equation.2. 7.html. the storage in Newton-GMRES is used in the inner iteration. Broyden’s method for the H-equation. Even so.9999. BROYDEN’S METHOD 129 For c = .3. . One should take some care in drawing general conclusions. while that in Broyden in the nonlinear iteration. 7. both implementations performed better than Newton-GMRES.ec-securehost. it may be the case Buy this book from SIAM at http://www. This electronic version is for personal use and may not be duplicated or distributed.9. c = .9999. 10 0 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 10 -6 10 -7 0 2 4 6 8 10 Iterations 12 14 16 18 Fig. c = . While Broyden’s method on this example performed better than Newton-GMRES. If many nonlinear iterations are needed. the restarted iteration has much more trouble.

2) of 2.4. At most five inner GMRES iterations were required. This is not supported by the local convergence theory for nonlinear problems. and τr = τa = h2 .ec-securehost.3 we will return to this example.com/SIAM/FR16. We preconditioned with the Poisson solver fish2d. We allowed for an increase in the residual. In § 8. Hence restarting the Broyden iteration every 8 iterations most closely corresponds to the Newton-GMRES requirements.6 million floating-point operations. u0 = 0. however it can be quite successful in practice. .4. This electronic version is for personal use and may not be duplicated or distributed. 7.4 we compare Broyden’s method without restarts (solid line) to Broyden’s method with restarts every 8 iterates (dashed line). In Fig.4. requires storage of n + 2 vectors {sj }n−1 . 130 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS that Broyden’s method may need restarts while the GMRES iteration for the linear problem for the Newton step may not. Convection-diffusion equation.3. Therefore the GMRES inner iteration for the Newton-GMRES approach needed storage for at most 10 vectors (the Krylov basis. We consider the nonlinear convectiondiffusion equation from § 6. which happened on the second iterate. We set C = 20. and z (when stored in the j=0 same place as F (x)). This approach Buy this book from SIAM at http://www. the residual norms do not monotonically decrease in the Broyden iteration (we fix this in § 8. F (xc + hv)) to accommodate the Newton-GMRES iteration. xc . F (xc ).2) and more storage was used.2. 7. as one would if the problem were truly linear.2 million floating-point operations.Copyright ©1995 by the Society for Industrial and Applied Mathematics. x.html. The unrestarted Broyden iteration converges in 12 iterations at a cost of roughly 2. However. The Broyden iteration when restarted every n iterates. Broyden’s method for nonlinear PDE. 10 1 10 Relative Nonlinear Residual 0 10 -1 10 -2 10 -3 10 -4 0 2 4 6 8 Iterations 10 12 14 16 Fig. We use the same mesh width h = 1/32 on a uniform square grid. This is a modest improvement over the NewtonGMRES cost (reported in § 6. One can do this in brsol by setting the third entry in the parameter list to 1.4. s.

html. This electronic version is for personal use and may not be duplicated or distributed.4 million floating-point operations.Copyright ©1995 by the Society for Industrial and Applied Mathematics.com/SIAM/FR16. but the convergence was extremely irregular after the restart. . BROYDEN’S METHOD 131 took 15 iterations to terminate at a cost of 2.ec-securehost. Buy this book from SIAM at http://www.

1) with homogeneous Dirichlet boundary conditions.4.4.2.5. See [102] for further generalizations.8.5. 7.6. V is M × N . Precondition with the Poisson solver fish2d.7.1 by showing that if A is a nonsingular N × N matrix.com/SIAM/FR16. This electronic version is for personal use and may not be duplicated or distributed. Modify nsol to use Broyden’s method instead of the chord method after the initial Jacobian evaluation. Set τa = τr = 10−8 .5.3.4 with Broyden’s method using various values for the parameter C. If this is the case. Experiment with various mesh widths. Extend Proposition 7.5. 7.5.9.45) Buy this book from SIAM at http://www. .5.5. 7. Solve the H-equation with Broyden’s method and c = 1. This is the Sherman–Morrison–Woodbury formula [68]. 1) × (0.3. 132 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 7.3.2.html.ec-securehost. then (A + U V )−1 = A−1 − A−1 U (I + V A−1 U )−1 V A−1 . 7. 7. Vary the parameters in the equations and the number of vectors stored by Broyden’s method and report on their effects on performance.1. Exercises on Broyden’s method 7. What happens in the differential equation examples if preconditioning is omitted? 7. How does the performance of Broyden’s method differ from Newton-GMRES? 7.5.1. What q-factor for linear convergence do you see? Can you account for this by applying the secant method to the equation x2 = 0? See [53] for a complete explanation. then A + U V is nonsingular if and only if I + V A−1 U is a nonsingular M × M matrix.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Duplicate the results reported in § 7. initial iterates. U is N × M . How does this compare with nsol on the examples in Chapter 5. Prove Proposition 7. Prove Proposition 7. so named because of its inferior performance in practice [63]. The bad Broyden method [26]. [199]. and values for the convection coefficient d. Compare the performance of Broyden’s method and Newton-GMRES on the Modified Bratu Problem −∇2 u + dux + eu = 0 on (0.1. Try to solve the nonlinear convection-diffusion equation in § 6.5.5.5. updates an approximation to the inverse of the Jacobian at the root so that B −1 ≈ F (x∗ )−1 satisfies the inverse secant equation −1 B+ y = s (7. 7.

Read about the algorithm in [172] and prove local superlinear convergence under the standard assumptions and the additional assumption that the sparsity pattern of B0 is the same as that of F (x∗ ). Implement the bad Broyden method. Does the relative performance of the methods change as h is decreased? This is interesting even in one space dimension where all matrices are tridiagonal. 7.5.11.5. and [112].10.7.4. the update is (y − Bc s)i sj (B+ )ij = (Bc )ij + i+1 2 k=i−1 sk for 1 ≤ i. and [189]. Note that only the subdiagonal. [92]. Discuss the differences in implementation from Broyden’s method. [125]. for example. This electronic version is for personal use and may not be duplicated or distributed. and compare the performance to that of Broyden’s method. How would the Schubert algorithm be affected by preconditioning? Compare your analysis with those in [158].html. and main diagonal are updated. are relevant to this exercise. y 2 2 Show that the bad Broyden method is locally superlinearly convergent if the standard assumptions hold. [27] is a quasi-Newton update for sparse matrices that enforces both the secant equation and the sparsity pattern. [93]. 7. BROYDEN’S METHOD 133 with the rank-one update −1 −1 B+ = Bc + −1 (s − Bc y)y T . Does anything go wrong? See [67] for more about this. apply it to the examples in § 7. References [101].5. j ≤ N and |i − j| ≤ 1.5.Copyright ©1995 by the Society for Industrial and Applied Mathematics.12.13. Implement the Schubert algorithm on an unpreconditioned discretized partial differential equation (such as the Bratu problem from Exercise 7.5. The Schubert or sparse Broyden algorithm [172].3.8) and compare it to Newton-GMRES and Broyden’s method.com/SIAM/FR16. 7. What can you conclude from your observations? Buy this book from SIAM at http://www. Use your result from Exercise 5. superdiagonal. 7.ec-securehost. For problems with tridiagonal Jacobian. Try to implement the bad Broyden method in a storage-efficient manner using the ideas from [67] and § 7. .14 in Chapter 5 to numerically estimate the q-order of Broyden’s method for some of the examples in this section (both linear and nonlinear).

.html.com/SIAM/FR16.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 134 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Buy this book from SIAM at http://www. This electronic version is for personal use and may not be duplicated or distributed.ec-securehost.

and because it is trivial to add a line search to an existing locally convergent implementation of Newton’s method.Copyright ©1995 by the Society for Industrial and Applied Mathematics. and continuation/homotopy methods [1]. The implementation and analysis of the line search method we develop in this chapter do not depend on whether iterative or direct methods are used to compute the Newton step or upon how or if the Jacobian is computed and stored. The reason for this is that f (x) = (1 + x2 )−1 is small for large x. [88]. but the length of s is too large. −1. We can see this effect in the sequence of iterates: 10.9 × 104 . where the algorithms can be motivated and intuition developed with a very simple example. can be used to accomplish the same objective. [70]. and hence the magnitude of the Newton step is much larger than that of the iterate itself.9 × 1017 . f (x) = arctan(x) ≈ ±π/2. [109]. 135 Buy this book from SIAM at http://www. [70]. [159]. a failure of Newton’s method that results from an inaccurate initial iterate. [25].com/SIAM/FR16. Of the many such algorithms we will focus on the class of line search methods and the Armijo rule [3]. .1. This electronic version is for personal use and may not be duplicated or distributed. we see that the Newton step s = −f (xc )/f (xc ) is pointing toward the root in the sense that sxc < 0. [154]. [25]. A prototype algorithm is given below.ec-securehost. [2]. If we look more closely. We select the line search paradigm because of its simplicity. [24].html. Other methods. 9. 2. [63]. such as trust region [153]. 8. implemented inexactly [24]. [173]. We begin with single equations. [198]. Chapter 8 Global Convergence By a globally convergent algorithm we mean an algorithm with the property that for any initial iterate the iteration either converges to a root of F or fails to do so in one of a small number of ways.5 × 109 . [181]. Single equations If we apply Newton’s method to find the root x∗ = 0 of the function f (x) = arctan(x) with initial iterate x0 = 10 we find that the initial iterate is too far from the root for the local convergence theory in Chapter 5 to be applicable. −138. This observation motivates the simple fix of reducing the size of the step until the size of the nonlinear residual is decreased. [129].

1. we obtain the sequence 10. This electronic version is for personal use and may not be duplicated or distributed. 8.html. 3. lines1(x. After the fourth iterate. 136 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Algorithm 8. We plot |arctan(x)/arctan(x0 )| as a function of the number of function evaluations with the solid line. (b) s = −f (x)/f (x) (search direction) (c) xt = x + s (trial point) (d) If |f (xt )| < |f (x)| then x = xt (accept the step) else s = s/2 goto 2c (reject the step) When we apply Algorithm lines1 to our simple example. however. some authors [63] refer to this as backtracking.4. 3.Copyright ©1995 by the Society for Industrial and Applied Mathematics. One can see that most of the work is in the initial phase of the iteration.56. The algorithm is to compute a search direction d which for us will be the Newton direction d = −f (xc )/f (xc ) Buy this book from SIAM at http://www.5. In the terminal phase. even in this terminal phase rapid convergence takes place only in the final three outer iterations.1. and 2 steplength reductions. We call this method a line search because we search along the line segment (xc . We plot the progress of the iteration in Fig. As we move from the right endpoint to the left.1.9. −6 × 10−10 .8. . 1.ec-securehost. 4. the minimum number of two function evaluations per iterate is required. where the local theory is valid. 9 × 10−4 .1. The quadratic convergence behavior of Newton’s method is only apparent very late in the iteration. decrease in the size of the nonlinear residual was obtained with a full Newton step. Iterates 1. −3. 2. To remedy this and to make analysis possible we replace the test for simple decrease with one for sufficient decrease.1.99. and 4 required 3. The outer iterations are identified with circles as in § 6. τ ) 1. 2.4. −0. −1. . xc − f (xc )/f (xc )) to find a decrease in |f (xc )|. However. One does not set x+ = x + s without testing the step to see if the absolute value of the nonlinear residual has been decreased.3. Do while |f (x)| > τr r0 + τa (a) If f (x) = 0 terminate with failure.com/SIAM/FR16. there is the possibility of oscillation about a solution without convergence [63]. f. In theory. −8. This is typical of the performance of the algorithms we discuss in this chapter. r0 = |f (x)| 2. Note the differences between Algorithm lines1 and Algorithm nsol.2. −. Algorithm lines1 almost always works well.

until f (x + s) satisfies (8. Do while |f (x)| > τr r0 + τa (a) If f (x) = 0 terminate with failure. A factor of Buy this book from SIAM at http://www.Copyright ©1995 by the Society for Industrial and Applied Mathematics.1) as easy as possible to satisfy. number intended to make (8. If |f (xt )| < (1 − αλ)|f (x)| then x = xt (accept the step) else λ = λ/2 goto 2(c)i (reject the step) This is essentially the algorithm implemented in the MATLAB code nsola. 1) is a small. The parameter α ∈ (0. Note that we must now distinguish between the step and the Newton direction. and then test steps of the form s = λd. This electronic version is for personal use and may not be duplicated or distributed. We will analyze Algorithm nsola in § 8. . something we did not have to do in the earlier chapters.html. Once sufficient decrease has been obtained we accept the step s = λd. 8. but positive. We follow the recent optimization literature [63] and set α = 10−4 . GLOBAL CONVERGENCE 0 137 10 10 10 Relative Nonlinear Residual 10 10 10 10 10 10 10 10 -1 -2 -3 -4 -5 -6 -7 -8 -9 -10 0 5 10 15 20 Function Evaluations 25 30 35 Fig. The condition in (8. nsola1(x.1) |f (xc + λd)| < (1 − αλ)|f (xc )|.1.2.ec-securehost.com/SIAM/FR16. τ ) 1. from [3] (see also [88]) is called the Armijo rule. r0 = |f (x)| 2.1) is called sufficient decrease of |f |. This strategy. f.2. (b) d = −f (x)/f (x) (search direction) (c) λ = 1 i. Algorithm 8. Newton–Armijo iteration for inverse tangent. We remark here that there is nothing critical about the reduction factor of 2 in the line search. with λ = 2−j for some j ≥ 0. xt = x + λd (trial point) ii.1.

138 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 10 could well be better in situations in which small values of λ are needed for several consecutive steps (such as our example with the arctan function). This electronic version is for personal use and may not be duplicated or distributed.5. Safeguarding is an important aspect of many globally convergent methods and has been used for many years in the context of the polynomial interpolation algorithms we discuss in § 8. We write this for the Armijo sequence as (8. Taking full Newton steps ensures fast local convergence. Our algorithm nsola incorporates these ideas.3. On the other hand. [86].Copyright ©1995 by the Society for Industrial and Applied Mathematics. or even use the chord method provided that the resulting direction dn satisfies (8. nsola(x. In Exercise 8. We will return to the issue of reduction in λ in § 8. Algorithm 8.2) F (xn )dn + F (xn ) ≤ ηn F (xn ) . is to minimize an interpolating polynomial based on the previous trial steps. 1.2). where 0 < σ0 < σ1 < 1.9 you are asked to implement the Armijo rule in the context of Algorithm nsol.1 [54]. This could be quite important if the search direction were determined by an inexact Newton method or an approximation to Newton’s method such as the chord method or Broyden’s method. [63]. r0 = F (x) 2. Typical values of σ0 and σ1 are . While (8. Our second extension is to allow more flexibility in the reduction of λ. 8.1). Analysis of the Armijo rule In this section we extend the one-dimensional algorithm in two ways before proceeding with the analysis. F. [133].1. in fact. Buy this book from SIAM at http://www.3. we accept any direction that satisfies the inexact Newton condition (6. .com/SIAM/FR16. approximate the Newton step by a quasi-Newton method.html. the analysis here applies equally to methods that solve the equations for the Newton step directly (so ηn = 0).2) is motivated by the Newton-iterative paradigm. Taking as large a fraction as possible helps move the iteration into the terminal phase in which full steps may be taken and fast convergence expected. The danger is that the minimum may be too near zero to be of much use and. developed in § 8. reducing λ by too much can be costly as well.5. the iteration may stagnate as a result.1 and . First. η). The parameter σ0 safeguards against that.ec-securehost.3.2. terminate with failure. In that event a reduction of a factor of 8 would require three passes through the loop in nsola1 but only a single reduction by a factor of 10.2. [87]. τ. Do while F (x) > τr r0 + τa (a) Find d such that F (x)d + F (x) ≤ η F (x) If no such d can be found. We allow for any choice that produces a reduction that satisfies σ0 λold ≤ λnew ≤ σ1 λold . One such method.

a local minimum of F which is not a root. Note how the safeguard bound σ0 enters this result. F (xn ) is singular for some n. Clearly if F has no roots. xn → x. 1 − α) Buy this book from SIAM at http://www.1. Let {xn } be the iteration produced by Algorithm nsola with initial iterate x0 .2. η ] ⊂ (0. We show this. and F (x)−1 ≤ mf on the set Ω({xn }. This electronic version is for personal use and may not be duplicated or distributed. The inner iteration could terminate with a failure. σ1 ] λ = σλ goto 2(b)i Note that step 2a must allow for the possibility that F is ill-conditioned and that no direction can be found that satisfies (8. xt = x + λd ii.2). ill-conditioning of F might be detected as part of the factorization. xn → ∞. F is Lipschitz continuous with Lipschitz constant γ. as do [25] and [70]. and the convergence is q-quadratic. We begin with a formal statement of our assumption that F is uniformly Lipschitz continuous and bounded away from zero on the sequence {xn }. mf > 0 such that F is defined. full steps are taken in the terminal phase. step 2a were implemented with a direct factorization method.html. r) = ∞ {x | x − xn ≤ r}. then the iteration converges to a root of F at which the standard assumptions hold.com/SIAM/FR16. Let x0 ∈ RN and α ∈ (0. We will see that if F has no roots then either F (xn ) has a limit point which is singular or {xn } becomes unbounded. 2. Note also the very modest conditions on ηn . the algorithm must fail.2. There are r.ec-securehost. GLOBAL CONVERGENCE 139 (b) λ = 1 i. n=0 Assumption 8. If.1 implies that the line search phase (step 2b in Algorithm nsola) will terminate in a number of cycles that is independent of n. γ. Assume that {xn } is given by Algorithm nsola with (8. The convergence theorem is the remarkable statement that if {xn } and { F (xn )−1 } are bounded (thereby eliminating premature failure and local minima of F that are not roots) and F is Lipschitz continuous in a neighborhood of the entire sequence {xn }.1.Copyright ©1995 by the Society for Industrial and Applied Mathematics. by giving a uniform lower bound on λn (assuming that F (x0 ) = 0). If F (xt ) < (1 − αλ) F (x) then x = xt else Choose σ ∈ [σ0 . ¯ 3.2.3) ¯ {ηn } ⊂ (0. The algorithm can fail in some obvious ways: 1. for example. Assumption 8. . Lemma 8. 1) be given.

Then either F (x0 ) = 0 or (8.html. for any λ ∈ [0.ec-securehost. which completes the proof.2) we have. . ¯ Therefore. This electronic version is for personal use and may not be duplicated or distributed.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Theorem 8. by Assumption 8. Now we can show directly that the sequence of Armijo iterates either is unbounded or converges to a solution.2. ¯ ξn ≤ ηn F (xn ) ≤ η F (xn ) . by the bound ηn ≤ η . 140 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS and that Assumption 8. 1 0 ¯ m2 γ(1 + η )2 λ2 F (xn ) f 2 ¯ m2 γλ(1 + η )2 F (x0 ) f 2 2 ≤ (1 − λ + η λ) F (xn ) + λ F (xn ) ¯ < (1 − αλ) F (xn ) . ¯ The step acceptance condition in nsola implies that { F (xn ) } is a decreasing sequence and therefore dn = F (xn )−1 (ξn − F (xn )) ≤ mf (1 + η ) F (xn ) . ¯ This shows that λ can be no smaller than σ0 λ2 .1 holds.2. which will be implied by ¯ ¯ λ ≤ λ2 = min λ1 . Proof. mf F (x0 ) m2 F (x0 ) (1 + η )2 γ ¯ f . 1] F (xn + λdn ) = F (xn ) + λF (xn )dn + = (1 − λ)F (xn ) + λξn + where. By the fundamental theorem of calculus and (8.1 says even more. xn + λdn ] whenever ¯ λ ≤ λ1 = r/(mf F (x0 ) ). 2(1 − α − η ) ¯ (1 + η )2 m2 γ F (x0 ) ¯ f .4) ¯ λn ≥ λ = σ0 min r 2(1 − α − η ) ¯ . The Buy this book from SIAM at http://www. ¯ Lipschitz continuity of F implies that if λ ≤ λ1 then F (xn + λdn ) ≤ (1 − λ) F (xn ) + λ¯ F (xn ) + η ≤ (1 − λ) F (xn )| + λ¯ F (xn ) + η γ 2 λ dn 2 2 1 0 (F (xn + tλdn ) − F (xn ))λdn dt (F (xn + tλdn ) − F (xn ))λdn dt.2.1 F is Lipschitz continuous on the line segment [xn .com/SIAM/FR16. Let n ≥ 0.

Theorem 8. The iteration can be very costly if the Jacobian must be Buy this book from SIAM at http://www. Moreover. for example.2. Eventually |xnk − x∗ | < r. Theorem 6. The boundedness of the sequence {xn } implies that a subsequence {xnk } converges. Otherwise Lemma 8. where r is the radius in Assumption 8.1 implies that F (xn ) converges q-linearly to zero with q-factor at ¯ most (1 − αλ). Assume that {xn } is given by Algorithm nsola.1 holds. If F (xn ) = 0 for some n. then the storage requirements are roughly the same as for Algorithm nsolgm if one reuses the storage location for the finite difference directional derivative to test for sufficient decrease. full steps are taken for n sufficiently large.1 and therefore the standard assumptions hold at x∗ .1 implies that the standard assumptions hold at x∗ = xn .2 hold is (8.2.4.html. so in the terminal phase of the iteration the step lengths are all 1 and the convergence is q-quadratic. numerically determine if the chord direction satisfies the inexact Newton condition. . Implementation of the Armijo rule The MATLAB code nsola is a modification of nsolgm which provides for a choice of several Krylov methods for computing an inexact Newton direction and globalizes Newton’s method with the Armijo rule. the Newton iteration with x0 = x will remain in B(δ) and converge qquadratically to x∗ . will remain in B(δ) and converge to x∗ . Since the standard assumptions hold at x∗ . Since F is continuous. the entire sequence.2). If GMRES is used as the linear solver. via a forward difference approximation to F (xn )d. 1) be given. In Exercise 8. Then {xn } converges to a root x∗ of F at which the standard assumptions hold. say to x∗ . F (x∗ ) = 0. {ηn } satisfies (8.Copyright ©1995 by the Society for Industrial and Applied Mathematics. and that Assumption 8.1 is violated) or the iteration becomes unbounded. then we are done because Assumption 8. Let x0 ∈ RN and α ∈ (0. Thus far we have used for d the inexact Newton direction. in principle. 8.2).2. there is δ such that if x ∈ B(δ).2. Proof.5.1. In this case the storage requirements are roughly the same as for Algorithm nsol.2. This is all that one could ask for in terms of robustness of such an algorithm.9 you are asked to do this with nsol and.3).3. This electronic version is for personal use and may not be duplicated or distributed. which is easy to verify. All that one needs to make the theorems and proofs in § 8. At this stage we have shown that if the Armijo iteration fails to converge to a root either the continuity properties of F or nonsingularity of F break down as the iteration progresses (Assumption 8. {xn } is bounded. and the convergence behavior in the final phase of the iteration is that given by the local theory for inexact Newton methods (Theorem 6.ec-securehost.2.1. Hence as soon as xnk ∈ B(δ).2 applies and hence full steps can be taken. with each iteration.1. It could well be advantageous to use a chord or Broyden direction. GLOBAL CONVERGENCE 141 sequence converges to a root at which the standard assumptions hold. We use the l2 norm in that code and in the numerical results reported in § 8. not just the subsequence.com/SIAM/FR16.

. The second. Buy this book from SIAM at http://www. Polynomial line searches. . If one can model the decrease in F 2 as the steplength is reduced. . We can use this iteration history to model the scalar function f (λ) = F (xn + λdn ) 2 2 with a polynomial and use the minimum of that polynomial as the next steplength. 142 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS evaluated and factored many times during the phase of the iteration in which the approximate solution is far from the root. one might expect to be able to better estimate the appropriate reduction factor. We consider two topics in this section. f (0) = F (xn ) 2 is known. if λt > σ1 λc . otherwise. is again based on [67] and is more specialized.10.3. This electronic version is for personal use and may not be duplicated or distributed. how Broyden’s method must be implemented. F (xn + λ1 dn ) 2 . (8.ec-securehost. After λc has been rejected and a model polynomial computed.5. In practice such methods usually perform better than constant reduction factors. We consider two ways of doing this that use second degree polynomials which we compute using previously computed information. In this approach we use values of f and f at λ = 0 and the value of f at the current value of λ to construct a 2nd degree interpolating polynomial for f .Copyright ©1995 by the Society for Industrial and Applied Mathematics.1. .com/SIAM/FR16. we compute the minimum λt of that polynomial analytically and set   σ0 λc            if λt < σ0 λc . If we have rejected k steps we have in hand the values F (xn ) 2 . Evaluation of F (xc )dc if the full step is not accepted (here dc is the Broyden direction) would not only verify the inexact Newton condition but provide enough information for a two-point parabolic line search. say). F (xn + λk−1 dn ) 2 .6) f (0) = 2(F (xn )T dn )T F (xn ) = 2F (xn )T (F (xn )dn ). 8. The reader is asked to pursue this in Exercise 8. is applicable to all methods. Clearly f (0) can be computed as 2 (8. but might respond well to a more aggressive steplength reduction (by factors of 1/10. polynomial interpolation as a way to choose λ.html. In this section we consider an elaboration of the simple line search scheme of reduction of the steplength by a fixed factor. Our Broyden–Armijo code does not explicitly check the inexact Newton condition and thereby saves an evaluation of F . The first. The motivation for this is that some problems respond well to one or two reductions in the steplength by modest amounts (such as 1/2) and others require many such reductions.5) λ+ = σ 1 λc λt Two-point parabolic model.

1. σ1 λ] or reject the parabolic model and simply set λ+ to σ0 λc or σ1 λc . where c= f (λc ) − f (0) − f (0)λc . f (λc ). then c > 0 and p(λ) has a minimum at λt = −f (0)/(2c) > 0. and f (λ− ). λc . GLOBAL CONVERGENCE 143 Use of (8. p agree with f. . then we set λt to the minimum of p λt = −p (0)/p (0) and apply the safeguarding step (8. An alternative to the two-point model that avoids the need to approximate f (0) is a three-point model. where λc and λ− are the most recently rejected values of λ. In any case. we implement this method with σ0 = . Buy this book from SIAM at http://www. Three-point parabolic model. f at 0 and p(λc ) = f (λc ) is p(λ) = f (0) + f (0)λ + cλ2 . We take the latter approach and set λ+ = σ1 λc . The polynomial p(λ) such that p.com/SIAM/FR16. If it is not. In the MATLAB code nsola from the collection. The MATLAB code parab3p implements the three-point parabolic line search and is called by the two codes nsola and brsola. we set λ = σ1 and try again. it may be necessary to compute a new search direction. We then compute λ+ with (8.html. This is a possibility with directions other than inexact Newton directions.Copyright ©1995 by the Society for Industrial and Applied Mathematics. If the full step is rejected. After the second step is rejected.6) requires evaluation of F (xn )dn .5) to compute λ+ . we have the values f (0).ec-securehost. which uses f (0) and the two most recently rejected steps to create the parabola. such as the Broyden direction.5). If p (0) ≤ 0 one could either set λ+ to be the minimum of p on the interval [σ0 λ. σ1 = . This electronic version is for personal use and may not be duplicated or distributed. λ2 c Our approximation of f (0) < 0. λc λ− We must consider two situations. In this approach one evaluates f (0) and f (1) as before.5. our approximation to f (0) should be negative. so if f (λc ) > f (0). The polynomial that interpolates f at 0. which can be obtained by examination of the final residual from GMRES or by expending an additional function evaluation to compute F (xn )dn with a forward difference. λ− is p(λ) = f (0) + λ λc − λ− (λ − λ− )(f (λc ) − f (0)) (λc − λ)(f (λ− ) − f (0)) + . If p (0) = 2 (λ− (f (λc ) − f (0)) − λc (f (λ− ) − f (0))) λc λ− (λc − λ− ) is positive.

3.3 is that the simple relation between the sequence {wn } and {vn } is changed.8) and (8. yn − Bn sn sn −1 T −1 .44) (8.5.2. the polynomial modeling approach is heuristic.3.7) yn − Bn sn = F (xn+1 ) − (1 − λn )F (xn ).8) to construct Bc F (x+ ) as we did in Algorithm brsol. 8.3.41) when λn = 1 and λn+1 = 1 (so dn+1 = sn+1 ). using the notation in § 7. we set un = we can use (8. sn 2 sn 2 If.ec-securehost.8) reduces to (7. The safeguarding and Theorem 8.5. The relation y − Bc s = F (x+ ) was not critical to this and we may also incorporate a line search at a cost of a bit of complexity in the implementation. Note that (8. our approach is a nonlinear version of the method in [67]. [63].7) and (7. 144 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS Interpolation with a cubic polynomial has also been suggested [87].8) = λn where −dn+1 sn + (λ−1 − 1) n sn 2 sn 2 −sn+1 /λn+1 sn + (λ−1 − 1) n sn 2 sn −1 dn+1 = −Bn+1 F (xn+1 ) 2 . .9) dn+1 = − −1 −1 sn 2 Bn F (xn+1 ) − (1 − λn )sT Bn F (xn+1 )sn n 2 −1 sn 2 + λn sT Bn F (xn+1 ) n 2 −1 We then use the second equality in (8.8) and the relation dn+1 = − I − wn sT n sn 2 −1 Bn F (xn+1 ) to solve for dn+1 and obtain an analog of (7. Buy this book from SIAM at http://www.9) is left to Exercise 8. vn = .com/SIAM/FR16. [86].Copyright ©1995 by the Society for Industrial and Applied Mathematics.1 provide insurance that progress will be made in reducing F . The difference from the development in § 7.html. Clearly. In Chapter 7 we implemented Broyden’s method at a storage cost of one vector for each iteration. [133]. This electronic version is for personal use and may not be duplicated or distributed. We can use the first equality in (8. As in § 7. Broyden’s method. is the search direction used to compute xn+1 . then −1 sn = −λn Bn F (xn ) and hence (8. Verification of (8. and wn = (Bn un )/(1 + vn Bn un ).2. If a line search is used.38) to obtain wn = λn (8.

The MATLAB implementation of brsola in the collection illustrates this point.3. Do while n ≤ maxit (a) n = n + 1 (b) x = x + sn (c) Evaluate F (x) (d) If F (x) (e) 2 ≤ τr r0 + τa exit.5.10 the reader is asked to fully develop and implement these ideas. the storage of the globalized methods exceeds that of the local method by one vector. As a side effect. provided in the collection of MATLAB codes. Examples for Newton–Armijo The MATLAB code nsola is a modification of nsolgm that incorporates the three-point parabolic line search from § 8. n − 1 a = −λj /λj+1 . uses a three-point parabolic line search and performs a restart when storage is exhausted like Algorithm brsol does. the storage requirements of Algorithm brsola would appear to be essentially the same as those of brsol. d = −F (x). τ.2).3. s0 = λ0 d 2. n = −1.2) before beginning the line search. This electronic version is for personal use and may not be duplicated or distributed. there is no guarantee that the Broyden direction is a direction of decrease for F 2 . z = −F (x) ii.ec-securehost.Copyright ©1995 by the Society for Industrial and Applied Mathematics. . compute λ0 .1 and also changes η using (6.com/SIAM/FR16. 8. brsola(x.1. f (0). store sn+1 2 − λn sT z) n and λn+1 . GLOBAL CONVERGENCE 145 Algorthm brsola uses these ideas. So. Algorithm 8.4. maxit. Since dn+1 and sn+1 can occupy the same location.html. In fact. The MATLAB code returns with an error message if more than a fixed number of reductions in the step length are needed. However. F. nmax) 1.20) Buy this book from SIAM at http://www. computation of the step length requires storage of both the current point x and the trial point x + λs before a step can be accepted. Another approach would be to compute F (xc )d numerically if a full step is rejected and then test (8. Initialize: Fc = F (x) r0 = F (x) 2 . an approximation of f (0) can be obtained at no additional cost and a parabolic line search based on f (0). The MATLAB implementation. i. b = 1 − λj z = z + (asj+1 + bsj )sT z/ sj j 2 2 2 2 (f) d = ( sn 2 z + (1 − λn )sT zsn )/( sn n 2 (g) Compute λn+1 with a line search. (h) Set sn+1 = λn+1 d. In Exercise 8. It is possible that the Broyden direction fails to satisfy (8. and f (1) can be used. We do not include details of the line search or provide for restarts or a limited memory implementation. for j = 0.

As before. which requires a single function evaluation for each outer iteration when full steps can be taken.2 and 7.4. Counts of function evaluations corresponding to outer iterations are√ indicated by circles. Here. We use x0 = 10 as the initial iterate with τr = τa = 10−8 . when both methods succeed.com/SIAM/FR16.html.2 we plot the iteration history corresponding to the parabolic line search with the solid line and that for the constant reduction with the dashed line.2. We consider a much more diffi- −∇2 u + Cu(ux + uy ) = f with the same right hand side f . because of the large value of C and resulting ill-conditioning.21) from § 6. This makes a globalization strategy critical (Try it without one!). Convection-diffusion equation. The constant reduction line search took three reductions in the first two outer iterations and two each in following two. is needed to obtain an accurate solution. using GMRES as the linear solver.4. We caution the reader now. we can use the two-point parabolic line search.20) with the dashed Buy this book from SIAM at http://www. We compare this strategy. 31 × 31 mesh and homogeneous Dirichlet boundary conditions on the unit square (0. the quasi-Newton method. This is a tighter tolerance that in § 6.1. with the constant η method as we did in § 6. We base our absolute error criterion on the norm · 2 / N as we did in § 6. Since we have easy access to analytic derivatives in this example.2 and. . This electronic version is for personal use and may not be duplicated or distributed. 1) that we considered before.4. In § 8. that if the initial iterate is far from the solution. in all the figures we plot F (xn ) 2 / F (x0 ) 2 against the number of function evaluations required by all line searches and inner and outer iterations to compute xn .9.3 we compare the Broyden–Armijo method with Newton-GMRES for those problems for which both methods were successful.4 with γ = . 146 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS once the iteration is near the root. 8. the parabolic line search takes only one reduction in the next three iterations and none thereafter.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 1) × (0.4. In Fig.2.ec-securehost. The parabolic line search required 7 outer iterates and 21 function evaluations in contrast with the constant reduction searches 11 outer iterates and 33 function evaluations.25 with the solid line and using the sequence given by (6. 8. an inexact Newton method such as Newton-GMRES can succeed in many case in which a quasi-Newton method can fail because the quasi-Newton direction may not be an inexact Newton direction. Inverse tangent function.4. we set C = 100. In Fig. We compare the two-point parabolic line search with the constant reduction search (σ0 = σ1 = .5) for the arctan function. However. However. and will repeat the caution later. both line searches take three steplength reductions.3 we show the progress of the iteration for the unpreconditioned equation. We set τr = τa = h2 /10. initial iterate u0 = 0. 8. cult problem.4. We take (6. however.4. 8. For this problem we plot the progress of the iteration using η = . may well be most efficient. In the first outer iteration.

8. line.Copyright ©1995 by the Society for Industrial and Applied Mathematics.9 and ηmax = . The two approaches to selection of the forcing term performed very similarly. In the computation we Buy this book from SIAM at http://www. The iterations required 75.25.2.3. C = 100. This problem is very poorly conditioned and much tighter control on the inner iteration is needed than for the other problems. . 8. GLOBAL CONVERGENCE 0 147 10 10 Relative Nonlinear Residual -2 10 -4 10 -6 10 -8 10 -10 10 -12 0 5 10 15 20 Function Evaluations 25 30 35 Fig. Newton–Armijo for the PDE.3 (constant η) and 75.20) to γ = . We consider the preconditioned problem in Fig. We set the parameters in (6.html.4. 8.com/SIAM/FR16. Newton–Armijo for the arctan function. 10 0 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 0 100 200 300 400 500 Function Evaluations 600 700 800 Fig. This electronic version is for personal use and may not be duplicated or distributed. 759 (constant η) and 744 (varying η) function evaluations. and 25 (constant η) and 22 (varying η) outer iterations.4 (varying η) million floating point operations.ec-securehost.

We found that the Broyden–Armijo line search failed on all the unpreconditioned convection diffusion equation examples. When compared to the Broyden’s method results in § 6. and roughly 13.4. when increases in the residual were allowed. terminating in 4 outer iterations. The most efficient iteration. 9 outer iterations.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 8. As reported in § 6. We set the parameters in (6.21) with the fast Poisson solver fish2d.4.20) with the dashed line.2. For this problem we plot the progress of the iteration using η = .3. and 2. The line search in this case reduced the step length three times on the first iteration and twice on the second and third.4 we show the progress of the iteration for the preconditioned equation. Newton–Armijo for the PDE.html.20). 9 outer iterations. the Broyden–Armijo costs reflect Buy this book from SIAM at http://www.3 million floating-point operations. 8. and 12.25 with the solid line and using the sequence given by (6.4. In these cases the Jacobian is not well approximated by a low rank perturbation of the identity [112] and the ability of the inexact Newton iteration to find a good search direction was the important factor.99. Broyden–Armijo. We begin by revisiting the example in § 6. .6 million floating-point operations. 16 function evaluations. C = 100. required at most 16 inner iterations while the constant η approach needed at most 10. 8. In Fig. with the forcing term given by (6.2 with C = 20. The constant η iteration terminated after 79 function evaluations. The line search in the non-constant η case reduced the step length three times on the first iteration and once on the second.9 and ηmax = . This electronic version is for personal use and may not be duplicated or distributed. The iteration in which {ηn } is given by (6. 148 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS preconditioned (6.20) to γ = .2.4.4.com/SIAM/FR16.20) terminated after 70 function evaluations. 10 0 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 0 10 20 30 40 50 Function Evaluations 60 70 80 Fig. the Newton-GMRES iteration always took full steps.5 million floating-point operations.ec-securehost.

Figure 8.4. 8. needing to store 21 vectors. and 85 function evaluations.8 million floating-point operations while the Broyden–Armijo took 9 iterations and required 2 million floating-point operations.5. ηmax = . the performance of the two methods is more similar. roughly 16 million floating-point operations. . 42 outer iterates. GLOBAL CONVERGENCE 149 an improvement in efficiency. In our implementation of brsola the three-point parabolic line search was used. 8. and γ = .9.9 as we did in § 6.ec-securehost. The Broyden iteration in § 6. Newton-GMRES and Broyden–Armijo for the PDE.com/SIAM/FR16. For the more difficult problem with C = 100. and 123 function evaluations. and γ = .5.5 compares the Broyden–Armijo iteration (solid line) to the GMRES iteration (dashed line). We set τa = τr = h2 . The restarted Broyden–Armijo iteration took 19. This electronic version is for personal use and may not be duplicated or distributed. The steplength was reduced on the second iteration (twice) and the third (once).html. ηmax = .4. and therefore was much more efficient in terms of storage. the Broyden–Armijo iteration required storage for 37 vectors while the Newton-GMRES iteration needed at most 16 inner iterations.5 million floating-point operations. 8. C = 20.6 we compare our implementation of Newton–GMRES–Armijo (dashed line) to Broyden–Armijo (solid line).7.99. Restarting the Broyden iteration every 19 iterations would equalize the storage requirements with Newton–GMRES–Armijo. In Fig. Broyden’s method suffers under these conditions as one can see from the comparison of Broyden– Armijo (solid line) and Newton–GMRES–Armijo (dashed line) in Fig. 10 0 Relative Nonlinear Residual 10 -1 10 -2 10 -3 0 2 4 6 8 10 Function Evaluations 12 14 16 Fig. We set τa = τr = h2 /10.2 took 12 iterations and 2. In terms of storage. From these examples we see that Broyden–Armijo can be very efficient provided only a small number of iterations are needed. The Broyden–Armijo iteration required 34 nonlinear iterations.2.Copyright ©1995 by the Society for Industrial and Applied Mathematics. The storage requirements increase as the nonlinear iteration progresses and this fact puts Broyden’s Buy this book from SIAM at http://www.

Newton-GMRES and Broyden–Armijo for the PDE. 8. Newton-GMRES and restarted Broyden–Armijo for the PDE.html. This electronic version is for personal use and may not be duplicated or distributed. 150 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 0 10 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 0 10 20 30 40 50 60 Function Evaluations 70 80 90 Fig. C = 100. .6.com/SIAM/FR16. method at a disadvantage to Newton-GMRES when the number of nonlinear iterations is large. 8. Buy this book from SIAM at http://www.ec-securehost.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 10 0 10 Relative Nonlinear Residual -1 10 -2 10 -3 10 -4 10 -5 0 20 40 60 80 Function Evaluations 100 120 140 Fig.7. C = 100.

Show that if F is Lipschitz continuous and the iteration {xn } produced by Algorithm nsola converges to x∗ with F (x∗ ) = 0. Explain your results. Apply your code to the following functions with x0 = 10.7.5.2.5.5.5.6. 8. How does the method proposed in [10] differ from the one implemented in nsola? What could be advantages and disadvantages of that approach? 8.. f (x) = x(1 + sin(1/x)) 5. Modify brsola to test the Broyden direction for the descent property and use a two-point parabolic line search. What could you do if the Broyden direction is not a descent direction? Apply your code to the examples in § 8. He puts $10. If the monthly payments are $713. f (x) = arctan(x) (i. and TFQMR.000. This electronic version is for personal use and may not be duplicated or distributed.9).) 2.html. Try to do it in such a way that the chord direction is used whenever possible. Use nsola to duplicate the results in § 8.e. Vary the convection coefficient in the convection-diffusion equation and the mesh size and report the results.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Experiment with other linear solvers such as GMRES(m). Duplicate the results in § 8. Verify (8. f (x) = . Implement the Newton–Armijo method for single equations.10.4.5.2.3.8) and (8.5. Modify nsol.5.1.5.5. f (x) = 2 + sin(x) 7.4. f (x) = 1 + x2 8. A numerical analyst buys a German sports car for $50.5. Bi-CGSTAB.000 down and takes a 7 year installment loan to pay the balance. 8.40.9. to use the Armijo rule. the hybrid Newton algorithm from Chapter 5.8. 8. what is the interest rate? Assume monthly compounding. This is interesting in the locally convergent case as well. Exercises on global convergence 8. You might use the MATLAB code nsola to do this. 8.9 + arctan(x) 4.1. f (x) = ex 6. 8.ec-securehost. then F (x∗ ) is singular. GLOBAL CONVERGENCE 151 8. In what ways are the results in [25] and [70] more general than those in this section? What ideas do these papers share with the analysis in this section and with [3] and [88]? 8.com/SIAM/FR16. f (x) = arctan(x2 ) 3. Buy this book from SIAM at http://www.5. 8.5. .4. 1.

ec-securehost. . 8. Does the secant method for equations in one variable always give a direction that satisfies (8.11.2.4.1 is applicable? 8.12. when does it? How would you implement a secant-Armijo method in such a way that the convergence theorem 8. 152 ITERATIVE METHODS FOR LINEAR AND NONLINEAR EQUATIONS 8.2) with ηn bounded away from 1? If not.5. This electronic version is for personal use and may not be duplicated or distributed.5.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Compare the results with the examples in § 8. Modify nsola to use a cubic polynomial and constant reduction line searches instead of the quadratic polynomial line search.13.html. Under what conditions will the iteration given by nsola converge to a root x∗ that is independent of the initial iterate? Buy this book from SIAM at http://www.com/SIAM/FR16.5.

1990. 22 (1951).. 37 (1981). R.. Philadelphia. S. 126 (1983). Manteuffel. Atkinson. John Wiley. Efficient methods to calculate Chandrasekhar’s H-functions. 16 (1966). J. [3] L. 17–31. Quart. Society for Industrial and Applied Mathematics. . Fund. ¨ [15] H. Global approximate Newton methods. [4] W. S. ´ ´ [13] N. Demmel. Donato. M. Georg. Numer. pp. Ann. 148–152. pp.. Allgower and K. Math. p. 279–295.. Numerical Continuation Methods : An Introduction. pp. Minimization of functions having Lipschitz-continuous first partial derivatives. [6] K. V. Campbell. E. pp.. This electronic version is for personal use and may not be duplicated or distributed. pp. [11] M. [2] . 17–29.. Bibliography [1] E. PA. The principle of minimized iterations in the solution of the matrix eigenvalue problem. Pacific J. F. Numer. [8] O.. Statist. Numerical Solution 153 Buy this book from SIAM at http://www. An inverse matrix adjustment arising in discriminant analysis. Brakhage. 107–111. 183–196.. Pozo. Cambridge. Petzold. Bosma and W. Brenan. Ashby. 2 (1960). SIAM J. Math. Dongarra. Simplicial and continuation methods for approximating fixed points and solutions to systems of equations. Engrg. A. Cambridge University Press. and H. e e [10] R. Bank and D. 14 (1979). [12] R. [5] S. 1989. Internat. Math. Math. J. Sur les op´rations dans les ensembles abstraits et leur applications e aux ´quations int´grales. T. E. Appl. Chan. Armijo.. 1–3. 283. 2nd. A comparison of adaptive Chebyshev and least squares polynomial preconditioning for Hermetian positive definite linear systems. C.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Eijkhout. T. 22 (1973). H. L. 133–181. New York. Methods. Uber die numerische Behandlung von Integralgleichungen nach der Quadraturformelmethode. SIAM Rev. [16] K. Statist. Numer. A. [7] . 22 (1980). Johnson.com/SIAM/FR16.ec-securehost. Comput. Barlett. Bicanic and K. Astrophys. pp. Rose. pp. pp. Romine. [14] P. Springer-Verlag. Math.. E. Banach. S. Astron. Otto. 3 (1922). Iterative variants of the Nystr¨m method for the numerical o solution of integral equations. Numer. B. 9 (1951). 28–85. ed. Math.. pp. R. Arnoldi. Who was ‘-Raphson’?. F. Axelsson. 1993. and J. E. Sci. DeRooij. J. pp. L. 13 (1992). New York. and L.html. Berry. van der Vorst. 1– 29. J. 1994. [9] S. J. Iterative Solution Methods. Templates for the Solution of Linear Systems: Building Blocks for Iterative Methods. An Introduction to Numerical Analysis. Barrett. Math.

This electronic version is for personal use and may not be duplicated or distributed. [35] R. Cai. Algorithms for Minimization Without Deriviatives. Hybrid Krylov methods for nonlinear systems of equations. 450–481. PA. in Proceedings of Internationaler Kongreß uber Anwendung der Mathematik in ¨ dem Ingenieurwissenschaften. J. Campbell. 24 (1987). [19] W. 577–593. 31 (1989). T. 154 BIBLIOGRAPHY of Initial Value Problems in Differential-Algebraic Equations. Schnabel. Numer. pp. 50. Brent. Philadelphia. 10 (1989). Using Krylov methods in the solution of large-scale differential-algebraic systems. N. Journal of Integral Equations and Applications. Philadelphia. Cambridge Univ. 19 (1965). Keyes. PhD thesis. SIAM J. D. Comput. 1987. L. 1467–1488. C. Tech. Ipsen.. 1038–1051.. Newton-KrylovSchwarz methods in CFD. pp. F. in Proceedings of the International Workshop on the Navier-Stokes Equations. Meyer. Applic. Cavanaugh. Vieweg Verlag. Press. University of Maryland. Brown. Kelley. 14 in Classics in Applied Mathematics. [29] W. The convergence of an algorithm for solving sparse nonlinear systems. [26] C. Byrne. 63 (1994). C. G. 40–91. [22] P. Difference Equations and Iterative Processes. Math. Dennis. Math. 11 (1990). N. J. A Multigrid Tutorial. 25 (1971). More. G. 129–156. D. Math. Tidriri. Inst. Comput. Tech... A local convergence theory for combined inexact–Newton/ finite– difference projection methods. 285–294. D. Kelley. D. pp. SIAM J. [17] R. [18] . Hindmarsh. [21] P. Sci. and A. 1960. R. Representation of quasiNewton matrices and their use in limited memory methods. Brown. C. Xue. Sci. BIT. Brown and A.com/SIAM/FR16. Optim. W.html. and Z. Braunschweig. Comput. Appl. March 1995. N. Programming. Anal. ed. E. [34] S. North Carolina State University.. Comput. D. Brown. [25] . Cambridge Tracts. no. Meyer. Zur Konvergenz einiger verfahren der konjugierten Richtungen. J. 10 (1973). to appear. C. Center for Research in Scientific Computation. and M. N. Cambridge. I. [33] S.. [23] P. Math. H.. July 1994. Report CRSC-TR95-13. Report CRSC-TR94-10. SIAM J. 1970.. Englewood Cliffs. 4 (1994). Buy this book from SIAM at http://www. C. Gropp. 12 (1973). Petzold. Byrd. 297–330. I. Society for Industrial and Applied Mathematics. Statist. 1975. G. Reduced storage matrix methods in stiff ODE systems. SIAM J. SIAM J. W. Convergence estimates for solution of integral equations with GMRES. 407–434. [32] X. C. Rannacher. Broyden. Nocedal. R. 1996. J. pp. Convergence theory of nonlinear Newton-Krylov algorithms. pp. pp. C. Broyden. Statist. . pp. E. and C. [20] P. F. 223– 246. The Mathematics of Radiative Transfer.-C. Some efficient algorithms for solving systems of nonlinear equations. Busbridge. and R. Campbell.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Saad. J. On the local and superlinear convergence of quasi-Newton methods. and L. 327–344. T. [24] P. [27] . A class of methods for solving nonlinear simultaneous equations. N. 1973. Burmeister. 1994. Comput. Maths. and J. pp. VODE: A variable coefficient ode solver. SIAM. pp..ec-securehost.. B. Hindmarsh. Ipsen.. Prentice-Hall. Brown and Y. [30] I. NJ. ´ [28] C. Weimar. Anal. D. Sci. L. Q. Hindmarsh. Briggs. C. Center for Research in Scientific Computation. Comput. No. A. Notes in Numerical Fluid Mechanics. 15 (1994). pp. Numer. pp. SIAM J. North Carolina State University. [31] R. GMRES and the minimal polynomial. to appear.

Powell.. Proceedings of the Fourth International Symposium on Domain Decomposition Methods. Dekker. Chandrasekhar. pp. T. More. [45] P. Newton’s method at singular points I. 187–209.. 42 (1983). in Constructive Aspects of the Fundamental Theorem of Algebra. Numer. P. [55] R. Philadelphia. Appl. W. G. Golub. [41] S. B.. pp. H. Meurant. Math. Math. No. Decker and C. pp. USSR. [39] . Moscow. Gallopoulos. J. 219–229. eds. [40] . T. G. SIAM J. 10–26. Numer. 1988. 147–154. 1989. pp. eds. Sci. Eisenstat. [46] E. 309–332. eds. pp. 6 (1985).. ´ [42] T. Handbook for Matrix Computations. 15 (1994). Houston. Dembo and T. J. pp. 4. R.. ed. R.Proceedings of the Second International Symposium on Domain Decomposition Methods. 220–252. H. Philadelphia. 34 (1955). Daniel. Anal. PA. SIAM J. [43] T. Sci. 20 (1983). Anal. pp. This electronic version is for personal use and may not be duplicated or distributed. Academic Press. Radiative Transfer. Numer. E.. Numer.. 4 (1967). PA. Concus. Reid. . Steihaug.. Sublinear convergence of the chord method at singular points. V. Report TR/PA/94/18. [51] . Anal. F. SIAM J. ´ [38] T. SIAM J. 19 (1982). pp. Society for Industrial and Applied Mathematics. Tech. Society for Industrial and Applied Mathematics. S. Anal. Statist. in Sparse Matrix Computations. January 14–16. Glowinski.. 20 (1983). 13 (1974). K. Truncated Newton algorithms for large-scale optimization. Phys. J. Estimation of sparse Jacobian matrices and graph coloring problems. Craig. 296–314. Coleman and C. Philadelphia.com/SIAM/FR16. Convergence rates for Newton’s method at singular points. 1990. pp. [47] A.. BIBLIOGRAPHY 155 [36] F. Society for Industrial and Applied Mathematics. CA. W. H. Szeto. TX. J. 19 (1982). pp. Numer. Tong. Widlund. Frontiers in Appl. Kelley. VanLoan.. A quasiminimal residual variant of the Bi-CGSTAB algorithm for nonsymmetric systems. 64– 73. 66–70. W. Comput. Block preconditioning for the conjugate gradient method. and T.ec-securehost. Inexact Newton methods. Convergence acceleration for Newton’s method at singular points. pp. Steihaug. Chan. Numer. 117–119. F. [49] D. T. Philadelphia. 1976. M. 338.. Domain Decomposition Methods. [56] R. Programming. SIAM J. Coleman and J. SIAM J. The conjugate gradient method for linear and nonlinear operator equations. and G. 1960. Los Angeles. Anal. PA. 37–48. Domain Decomposition Methods. J. and O. pp.. [52] . pp. J. Chan. Domain Decomposition Methods. CERFACS. and D. Buy this book from SIAM at http://www. Proceedings of the Third International Symposium on Domain Decomposition Methods. Dembo. [44] P. [48] J. O’Leary. 26 (1983). Keller. 190–212. D. Inst. Math. 1990. and C. J. Bunch and D. eds.Copyright ©1995 by the Society for Industrial and Applied Mathematics. Concus. PA. Math. Numer. 400–408. Numer. Math. Rose. J. and J. New York. 1989. Chaitin-Chatelin. 1988. [50] D.. J.html. 22 (1985). SIAM J. December 1994. Society for Industrial and Applied Mathematics. Dover. Simoncini. Decker. 1991. 17 (1980). Golub. On the estimation of sparse Jacobian matrices. Comput.. and C. [37] T. Curtis. Finding a zero by means of successive linear interpolation. The N-step iteration procedures. pp. Kelley. R. [53] .. Is nonnormality a serious difficulty ?. Anal.. SIAM J. p. P. Broyden’s method for a class of problems having singular Jacobian at the root.. A generalized conjugate gradient method for the numerical solution of elliptic partial differential equations. SIAM J.. Periaux. 566–574. 1969. Anal. [54] T. Henrici.

. Berlin. C. Comput. Some devices for the solution of large sets of simultaneous linear equations (with an appendix on the reciprocation of partitioned matrices). A hybrid Chebyshev-Krylov subspace algorithm for solving nonsymmetric systems of linear equations. B.. 1984. 358–382. C. Eisenstat and H. 17 (1996). Optim. Numerical Methods for Unconstrained Optimization and Nonlinear Equations. Manteuffel. Convergence theorems for least change secant update methods. 21 (1984).Copyright ©1995 by the Society for Industrial and Applied Mathematics. Anal. Sci. J. Programming. pp. Inaccuracy in quasi-Newton methods: Local improvement theorems. Anal. 1982. Elman. 16–32. J. pp. Walker. SIAM J. Conjugate gradient methods for indefinite systems. Elman. Anal. [74] V. 427–459. Choosing the forcing terms in an inexact Newton method. Y. Amsterdam.. SIAM Rev. Numer. Quasi-Newton methods. 73–89. J. Martinez. pp. pp. in Numerical Analysis Dundee 1975. Statist. ed. Numer. and K. Academic Press. 1971. The London. Optim. Methods Engrg. E.. Philadelphia. and A. New York. Freund. R. SIAM J. L. E. North–Holland. This electronic version is for personal use and may not be duplicated or distributed.. pp. 18 (1981). Nonsymmetric Systems of Linear Equations. 4 (1994). and Dublin Philosophical Magazine and Journal of Science. [72] H. E. methods. pp. 46–89. 156 BIBLIOGRAPHY [57] J. Numer. Dennis. B. 17 (1981). 4 (1994). [61] . SIAM J.html. Schnabel. SIAM J. Dennis and J. [59] J. E. The application of quasiNewton methods in fluid mechanics. [66] . Iterative Methods for Large. [71] H. Deuflhard. Dennis and H. Internat.com/SIAM/FR16. 425–472. Toward a unified convergence theory for Newton-like methods. Sci. 28 (1974). pp. E. 244–276. Saad. Zhang. 7 (1986). Feng. in Nonlinear Functional Analysis and Applications. Numer. 840–855. Walter. [70] . SIAM Rev. PhD thesis. Saylor. W. [73] M. Fletcher.ec-securehost.. motivation and theory. 949–987. M. G. SIAM J. ´ [60] J. D. F. C. Edinburgh. [75] D. Buy this book from SIAM at http://www. Watson. CT. 760–778. Dennis and R.. pp. Strang. SIAM J. Least change secant updates for quasiNewton methods. and P. [67] P. Frank. Least-change sparse secant updates with inaccurate secant conditions. Faber and T. [63] . Necessary and sufficient conditions for the existence of a conjugate gradient method. SIAM J. ed. Impact of Computing in Science and Engineering. 2 (1990). pp. pp. Fast secant methods for the iterative solution of large nonsymmetric linear systems. A. 35 (1944). [64] J. 443–459. SIAM. pp. Numer. Springer-Verlag. pp. G.. Sparse. Yale University. [65] . 6 (1969). pp. Duncan... no. . [69] S. 707–718. pp. 1976. Triangular decomposition methods for solving reducible nonlinear systems of equations. More. 352– 362. New York.. Schnabel. pp. Comput. [68] W. [76] R. Globally convergent inexact Newton methods. J. [62] J. and X. 1996. 660–670. A characterization of superlinear convergence and its application to quasi-Newton methods. 22 (1985). 393–422. pp. 19 (1977). E. 16 in Classics in Applied Mathematics. Math. SIAM J. 493–507. 70–85.. Math. in Mathematical Programming Study 22: Mathematical programming at Oberwolfach II.. New Haven. Local convergence analysis of tensor methods for nonlinear equations. B. pp. Anal. Dennis. pp. Bathe. 549–560. [58] . 62 (1993).. On the Kantorovich hypothesis for Newton’s method. Rall. J. Walker. F. and R. Engelman. 21 (1979). Seventh Series. Comput. P.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

BIBLIOGRAPHY

157

[77] R. W. Freund, A transpose-free quasi-minimal residual algorithm for nonHermitian linear systems, SIAM J. Sci. Comput., 14 (1993), pp. 470–482. [78] R. W. Freund, G. H. Golub, and N. M. Nachtigal, Iterative solution of linear systems, Acta Numerica, 1 (1991), pp. 57–100. [79] R. W. Freund, M. H. Gutknecht, and N. M. Nachtigal, An implementation of the look-ahead Lanczos algorithm for non-Hermitian matrices, SIAM J. Sci. Comput., 14 (1993), pp. 137–158. [80] R. W. Freund and N. M. Nachtigal, QMR: a quasi-minimal residual algorithm for non-hermitian linear systems, Numer. Math., 60 (1991), pp. 315–339. [81] , An implementation of the QMR method based on coupled two-term recurrences, SIAM J. Sci. Comput., 15 (1994), pp. 313–337. [82] D. M. Gay, Some convergence properties of Broyden’s method, SIAM J. Numer. Anal., 16 (1979), pp. 623–630. [83] C. W. Gear, Numerical Initial Value Problems in Ordinary Differential Equations, Prentice-Hall, Englewood Cliffs, NJ, 1971. [84] R. R. Gerber and F. T. Luk, A generalized Broyden’s method for solving simultaneous linear equations, SIAM J. Numer. Anal., 18 (1981), pp. 882–890. [85] P. E. Gill and W. Murray, Quasi-Newton methods for unconstrained optimization, J. I. M. A., 9 (1972), pp. 91–108. [86] , Safeguarded steplength algorithms for optimization using descent methods, Tech. Report NAC 37, National Physical Laboratory Report, Teddington, England, 1974. [87] P. E. Gill, W. Murray, and M. H. Wright, Practical Optimization, Academic Press, London, 1981. [88] A. A. Goldstein, Constructive Real Analysis, Harper and Row, New York, 1967. [89] G. H. Golub and C. G. VanLoan, Matrix Computations, Johns Hopkins University Press, Baltimore, 1983. [90] A. Griewank, Analysis and modification of Newton’s method at singularities, PhD thesis, Australian National University, 1981. [91] , Rates of convergence for secant methods on nonlinear problems in Hilbert space, in Numerical Analysis, Proceedings Guanajuato , Mexico 1984, Lecture Notes in Math., No, 1230, J. P. Hennart, ed., Springer-Verlag, Heidelberg, 1986, pp. 138– 157. [92] , The solution of boundary value problems by Broyden based secant methods, in Computational Techniques and Applications: CTAC 85, Proceedings of CTAC, Melbourne, August 1985, J. Noye and R. May, eds., North Holland, Amsterdam, 1986, pp. 309–321. [93] , On the iterative solution of differential and integral equations using secant updating techniques, in The State of the Art in Numerical Analysis, A. Iserles and M. J. D. Powell, eds., Clarendon Press, Oxford, 1987, pp. 299–324. [94] A. Griewank and G. F. Corliss, eds., Automatic Differentiation of Algorithms: Theory, Implementation, and Application, Society for Industrial and Applied Mathematics, Philadelphia, PA, 1991. [95] A. Griewank and P. L. Toint, Local convergence analysis for partitioned quasi-newton updates, Numer. Math., 39 (1982), pp. 429–448. [96] , Partitioned variable metric methods for large sparse optimization problems, Numer. Math., 39 (1982), pp. 119–137. [97] W. A. Gruver and E. Sachs, Algorithmic Methods In Optimal Control, Pitman, London, 1980.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

158

BIBLIOGRAPHY

[98] M. H. Gutknecht, Variants of BICGSTAB for matrices with complex spectrum, SIAM J. Sci. Comput., 14 (1993), pp. 1020–1033. [99] W. Hackbusch, Multi-Grid Methods and Applications, vol. 4 of Springer Series in Computational Mathematics, Springer-Verlag, New York, 1985. [100] , Multigrid methods of the second kind, in Multigrid Methods for Integral and Differential Equations, Oxford University Press, Oxford, 1985. [101] W. E. Hart and S. O. W. Soul, Quasi-Newton methods for discretized nonlinear boundary problems, Journal of the Institute of Applied Mathematics, 11 (1973), pp. 351–359. [102] H. V. Henderson and S. R. Searle, On deriving the inverse of a sum of matrices, SIAM Rev., 23 (1981), pp. 53–60. [103] M. R. Hestenes and E. Steifel, Methods of conjugate gradient for solving linear systems, J. of Res. Nat. Bureau Standards, 49 (1952), pp. 409–436. [104] D. M. Hwang and C. T. Kelley, Convergence of Broyden’s method in Banach spaces, SIAM J. Optim., 2 (1992), pp. 505–532. [105] E. Isaacson and H. B. Keller, Analysis of numerical methods, John Wiley, New York, 1966. [106] L. Kantorovich and G. Akilov, Functional Analysis, 2nd ed., Pergamon Press, New York, 1982. [107] L. V. Kantorovich, Functional analysis and applied mathematics, Uspehi Mat. Nauk., 3 (1948), pp. 89–185. translation by C. Benster as Nat. Bur. Standards Report 1509. Washington, D. C., 1952. [108] H. B. Keller, Newton’s method under mild differentiability conditions, J. Comput. Sys. Sci, 4 (1970), pp. 15–28. [109] , Lectures on Numerical Methods in Bifurcation Theory, Tata Institute of Fundamental Research, Lectures on Mathematics and Physics, Springer-Verlag, New York, 1987. [110] C. T. Kelley, Solution of the Chandrasekhar H-equation by Newton’s method, J. Math. Phys., 21 (1980), pp. 1625–1628. [111] , A fast multilevel algorithm for integral equations, SIAM J. Numer. Anal., 32 (1995), pp. 501–513. [112] C. T. Kelley and E. W. Sachs, A quasi-Newton method for elliptic boundary value problems, SIAM J. Numer. Anal., 24 (1987), pp. 516–531. [113] , A pointwise quasi-Newton method for unconstrained optimal control problems, Numer. Math., 55 (1989), pp. 159–176. [114] , Fast algorithms for compact fixed point problems with inexact function evaluations, SIAM J. Sci. Statist. Comput., 12 (1991), pp. 725–742. [115] , A new proof of superlinear convergence for Broyden’s method in Hilbert space, SIAM J. Optim., 1 (1991), pp. 146–150. [116] , Pointwise Broyden methods, SIAM J. Optim., 3 (1993), pp. 423–441. [117] , Multilevel algorithms for constrained compact fixed point problems, SIAM J. Sci. Comput., 15 (1994), pp. 645–667. [118] C. T. Kelley and R. Suresh, A new acceleration method for Newton’s method at singular points, SIAM J. Numer. Anal., 20 (1983), pp. 1001–1009. [119] C. T. Kelley and Z. Q. Xue, Inexact Newton methods for singular problems, Optimization Methods and Software, 2 (1993), pp. 249–267. [120] , GMRES and integral operators, SIAM J. Sci. Comput., 17 (1996), pp. 217– 226. [121] T. Kerkhoven and Y. Saad, On acceleration methods for coupled nonlinear elliptic systems, Numerische Mathematik, 60 (1992), pp. 525–548.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Copyright ©1995 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.

BIBLIOGRAPHY

159

[122] C. Lanczos, Solution of linear equations by minimized iterations, J. Res. Natl. Bur. Stand., 49 (1952), pp. 33–53. [123] T. A. Manteuffel, Adaptive procedure for estimating parameters for the nonsymmetric Tchebychev iteration, Numer. Math., 31 (1978), pp. 183–208. [124] T. A. Manteuffel and S. Parter, Preconditioning and boundary conditions, SIAM J. Numer. Anal., 27 (1990), pp. 656–694. [125] E. S. Marwil, Convergence results for Schubert’s method for solving sparse nonlinear equations, SIAM J. Numer. Anal., (1979), pp. 588–604. [126] S. McCormick, Multilevel Adaptive Methods for Partial Differential Equations, Society for Industrial and Applied Mathematics, Philadelphia, PA, 1989. [127] J. A. Meijerink and H. A. van der Vorst, An iterative solution method for linear systems of which the coefficient matrix is a symmetric M-matrix, Math. Comput., 31 (1977), pp. 148–162. [128] C. D. Meyer, Matrix Analysis and Applied Linear Algebra, forthcoming. ´ [129] J. J. More, Recent developments in algorithms and software for trust region methods, in Mathematical Programming: The State of the Art, A. Bachem, M. Gr¨schel, and B. Korte, eds., Springer-Verlag, Berlin, 1983, pp. 258–287. o ´ [130] J. J. More and J. A. Trangenstein, On the global convergence of Broyden’s method, Math. Comput., 30 (1976), pp. 523–540. [131] T. E. Mott, Newton’s method and multiple roots, Amer. Math. Monthly, 64 (1957), pp. 635–638. [132] T. W. Mullikin, Some probability distributions for neutron transport in a half space, J. Appl. Probab., 5 (1968), pp. 357–374. [133] W. Murray and M. L. Overton, Steplength algorithms for minimizing a class of nondifferentiable functions, Computing, 23 (1979), pp. 309–331. [134] N. M. Nachtigal, S. C. Reddy, and L. N. Trefethen, How fast are nonsymmetric matrix iterations?, SIAM J. Matrix Anal. Appl., 13 (1992), pp. 778– 795. [135] N. M. Nachtigal, L. Reichel, and L. N. Trefethen, A hybrid gmres algorithm for nonsymmetric linear systems, SIAM J. Matrix Anal. Appl., 13 (1992). [136] S. G. Nash, Preconditioning of truncated Newton methods, SIAM J. Sci. Statist. Comput., 6 (1985), pp. 599–616. [137] , Who was Raphson? (an answer). Electronic Posting to NA-Digest, v92n23, 1992. [138] J. L. Nazareth, Conjugate gradient methods less dependent on conjugacy, SIAM Rev., 28 (1986), pp. 501–512. [139] O. Nevanlinna, Convergence of Iterations for Linear Equations, Birkh¨user, a Basel, 1993. [140] I. Newton, The Mathematical Papers of Isaac Newton (7 volumes), D. T. Whiteside, ed., Cambridge University Press, Cambridge, 1967–1976. [141] B. Noble, Applied Linear Algebra, Prentice Hall, Englewood Cliffs, NJ, 1969. [142] J. Nocedal, Theory of algorithms for unconstrained optimization, Acta Numerica, 1 (1991), pp. 199–242. [143] D. P. O’Leary, Why Broyden’s nonsymmetric method terminates on linear equations, SIAM J. Optim., 4 (1995), pp. 231–235. [144] J. M. Ortega, Numerical Analysis A Second Course, Classics in Appl. Math., No. 3, Society for Industrial and Applied Mathematics, Philadelphia, PA, 1990. [145] J. M. Ortega and W. C. Rheinboldt, Iterative Solution of Nonlinear Equations in Several Variables, Academic Press, New York, 1970. [146] A. M. Ostrowski, Solution of Equations and Systems of Equations, Academic

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR16.html.

Statist. pp. 6 (1890). 7 (1986). J. New York. J. NATO Conf.. pp. J. Stepleman et al. ill-conditioned equations and Newton’s method. de Math. NJ. R. Appl. D. Parlett. 105–124. [167] Y.. 24 (1987). Computer Science Department. and S. Wilkinson. in Numerical Methods for Nonlinear Algebraic Equations. Statist. [152] E. Inverse iteration. K. Numerical Analysis of Parametrized Nonlinear Equations. ex nova infinitarum serierum doctrina. New York. 1690. eds. [164] . pp. Powell. O. [150] G. Math. Sci. (1993). Broyden’s method in Hilbert space. Comput. New York. L.com/SIAM/FR16. The Chebyshev Polynomials. Amsterdam. 1968. Peaceman and H.. pp. Anal. 325–328. [148] B. Saad. 1980. Schultz. John Wiley. Saad and M. Comput. Smolarski. 865–881. pp. [163] Y. Least squares solution of sparse systems of nonlinear equations by a modified Marquardt algorithm.. Rall.. 1970. SIAM Rev. Mangasarian. Practical use of polynomial preconditionings for the conjugate gradient method. Analysis aequationum universalis seu ad aequationes algebraicas resolvendas methodus generalis. Convergence of the Newton process to multiple solutions. SIAM J. [169] P. 15 (1978). Sci. J. North Holand. Implementation of an adaptive algorithm Buy this book from SIAM at http://www. Parlett. 1974. [155] L. SIAM J. This electronic version is for personal use and may not be duplicated or distributed. Rice. [153] M. 87–114. Numer. pp. deducta ac demonstrata. [168] E. 437–445. 29 (1979). 1983. pp.. N. 145–210.. 65–68. Comput. 1960. Sci. Real Analysis. S. SIAM J. 23–37.Copyright ©1995 by the Society for Industrial and Applied Mathematics. [151] L. [165] . Programming. [160] J. Rivlin. in Nonlinear Programming 2. [159] W. C. Raphson. pp. 155–169. New York. Gordon and Breach. 9 (1961). Report 92-38. W.. in Proc. London. pp.. The Symmetric Eigenvalue Problem. A. A flexible inner-outer preconditioned GMRES algorithm.. Amsterdam. R. A look-ahead Lanczos algorithm for unsymmetric matrices.. The numerical solution of parabolic and elliptic differential equations. pp. University of Minnesota. 20 (1966). Sachs. Liu. ser 4. [161] T. eds.. et expedita. Comput. 11 (1955). Indus.html. 1975. Math. 1973. pp. [166] . SIAM J. at Cambridge. Saylor and D. Experiments on Gram-Schmidt orthogonalization.. Comput. D. Petzold. R. Tech. Math. 35 (1986). Macmillan. 1992. 1986. Robinson. Numer. M. John Wiley. [149] D. pp. Rheinboldt. GMRES a generalized minimal residual algorithm for solving nonsymmetric linear systems. 44 (1985). 993–986. Original in British Library. New York. 461–469. E. [154] . ILUT: A dual threshold incomplete LU factorization. R. Englewood Cliffs. Math. R. 160 BIBLIOGRAPHY Press. [147] B. [156] J. pp. Math. Meyer. Soc.ec-securehost. Least squares polynomials in the complex plane and their use for solving nonsymmetric linear systems. A description of DASSL: a differential/algebraic system solver. New York. B. Numer. [158] J. pp. Taylor. 1–27. Royden. and Z. North Holland. R. On Newton’s method for singular problems. pp.. pp. 856–869. SIAM J. H. in Scientific Computing. Prentice Hall. N. 28–41. 339–360. Reddien. Convergence properties of a class of minimization algorithms. 71–82. Anal. Picard. 2nd ed. Reid. L. [162] H. Academic Press. Rachford. July 1972. A hybrid method for nonlinear equations. M´moire sur la th´orie des ´quations aux d´riv´es partielles et la e e e e e e m´thode des approximations successives. Peters and J. [157] G. C. 6 (1985). W. .

A family of trust-regionbased algorithms for unconstrained minimization with strong global convergence properties. [191] H. L. Fourier analysis and the FACR algorithm for the discrete solution of Poisson’s equation on a rectangle. Adjustment of an inverse matrix corresponding to a change in one element of a given matrix.. This electronic version is for personal use and may not be duplicated or distributed. 47–67.. C. SIAM J. 154/156 (1991). A. J. A modification of Newton’s method. L. 21 (1984).com/SIAM/FR16. Linear Algebrra Appl.. Anal. R. Modification of a quasi-Newton method for nonlinear equations with sparse Jacobian. Uber unendlich viele Algorithmen zur Auflosung der Gleichungen. SIAM J. 21 (1950). 133–138. Statist. [177] . E... 124–127. Varga. pp.. Sorensen. Statist. 1994. Ann. ACM Trans.. Applied Mathematics Series. Math. Traub. Math. 10 (1989). pp. 19 (1967). van der Vorst. pp. 815–843. A. U.. New York. pp.) [175] A. F. Math. Sci. Numer. [189] P. National Bureau of Standards. pp. Ann. Tensor methods for nonlinear equations. Comput. 27–30. [178] K. Kernel polynomials in linear algebra and their numerical applications. pp. Swarztrauber and R. Prentice Hall. Statist. Sci. SIAM J. [187] . Ann. 369–385.. 163–178. FL. [176] J. [182] G. 2 (1870). Numer. [170] R. [184] G. Iterative Methods for the Solution of Equations. 19 (1982). D. [186] P. 64 (1993). ¨ ¨ [171] E. and R. SIAM J. 755–774. 28 (1988). a fast Lanczos-type solver for nonsymmetric linear systems. pp. Schroder.. 317–365. 1–22. BIT. Fourth Edition. 1964. pp. Math. Comput. Math. 15 (1994). Zh. SIAM J. Software. H. Shamanskii. Stewart.. Statist. Mat. 954–961. Morrison. [185] E. pp. [172] L. [190] J.. Toint. [183] G. A hybrid Arnoldi-Faber iterative method for nonsymmetric systems of linear equations. 213–239. MATLAB Primer. Boca Raton. Comput. 615–646. W. pp. N. 5 (1979). 19 (1977). SIAM Rev. Newton’s method with a model trust region modification. S. Starke and R. [179] D. Anal. Numer.. pp. BIBLIOGRAPHY 161 for Richardson’s method.. Sci. Sherman. Englewood Cliffs. 36–52. B. Comput. 199–209. 31 (1977). K. Schnabel and P. Swarztrauber. NJ. Ukran. Numer.ec-securehost. Comput. Sci. 20 (1949). (In Russian. 49 (1958). 13 Buy this book from SIAM at http://www. The methods of cyclic reduction. Numer. Sherman and W. Approximate cyclic reduction for solving Poisson’s equation. SIAM J. pp. N. Statist. [188] P. Sweet. Smolarski and P. Comput. CRC Press. 409–426. Schubert. Schultz. 14 (1978). 1973. Math. Sonneveld. 621.. Bi-CGSTAB: A fast and smoothly converging variant to Bi-CG for the solution of nonsymmetric systems. Alternating direction preconditioning for nonsymmetric systems of linear equations. pp. Starke. Schnabel. 22 (1985).S. CGS. pp. SIAM J. p. [174] V. E. pp. Sigmon.html. Anal. pp. Introduction to matrix computations. Academic Press. Frank. pp. Adjustment of an inverse matrix corresponding to changes in the elements of a given column or a given row of the original matrix (abstract). 24 (1970). 352–364. [181] D. [173] G. Stiefel.. C. A. On sparse and symmetric matrix updating subject to a linear equation. pp. Math. 8 (1987). B. On Newton-iterative methods for the solution of systems of nonlinear equations. Saylor. Anal. [180] P. SIAM J.Copyright ©1995 by the Society for Industrial and Applied Mathematics. An optimum iterative method for solving any linear system with a square matrix.. . Algorithm 541: Efficient FORTRAN subprograms for the solution of elliptic partial differential equations. H. 490–501. Byrd.

M. 1950. Iterative Solution of Elliptic Systems and Applications to the Neutron Diffusion Equations of Reactor Physics. Billups. The effect of rounding errors on Newton-like methods. [199] M. Walker and L. Academic Press. Anal. [194] E. IMA J. A simpler GMRES. 815–825. Vuik. Applied Numer.. Math. 6 (1994). 48 (1993). Implementation of the GMRES method using Householder transformations. [202] L. [192] H. Inverting modified matrices. Matrix Iterative Analysis. pp. [195] H. Residual smoothing and peak/plateau behavior in Krylov subspace methods. Walker. T. Math. Zhou and H. pp. F. Englewood Cliffs. C. Zhou.html. pp. Memorandum Report 42. Numer. Prentice Hall. Morgan. Journal Comput. J. 571–581. Princeton NJ. pp. L. The superlinear convergence behaviour of GMRES. 1971. F. pp. Varga. Alg. [196] . SIAM J. 3 (1983). 15 (1994). van der Vorst and C. Buy this book from SIAM at http://www. Young.. Wachspress. pp. [201] T. [198] L. 297–312. P. This electronic version is for personal use and may not be duplicated or distributed. 327–341. Woodbury. pp. 13 (1987). New York. Residual smoothing techniques for iterative methods. NJ. Numer. 1962. Sci. Comput. ACM Trans.. Prentice Hall.com/SIAM/FR16. Sci.. Software. A. Statistical Research Group.ec-securehost. Appl. S. Ypma. 9 (1989). Statist.. Lin. Iterative Solution of Large Linear Systems. Watson. [200] D. and A.. 19 (1995). . Comput. 1966. Englewood Cliffs. F. A. [193] R. 631–644. Walker.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 162 BIBLIOGRAPHY (1992). J. Math. Appl. SIAM J. pp. S. NJ. 279–286. 281–310. Algorithm 652: HOMPACK: A suite of codes for globally convergent homotopy algorithms. [197] H. 109–118.

67 Convection-Diffusion Equation Broyden’s method. 47 Broyden sequence. This electronic version is for personal use and may not be duplicated or distributed. 47 Bi-Conjugate Gradient. 50 algorithm. 50 finite difference. 14 Conjugate Gradient Squared. 67 CGNE. . 130 Newton–Armijo. 49 Conjugate transpose. 74 convergence. 118. 107 Characteristic polynomial.ec-securehost. 145 163 Cauchy sequence. 128 Newton’s method. 120 implementation. 74 algorithm. 76 Condition number. 119 Buy this book from SIAM at http://www. 47 algorithm. 133 Brsol algorithm. 130 Convergence theorem Broyden’s method linear equations. 77 Armijo rule. 47 Bi-orthogonality. 87 solution Broyden’s method. 47 Bi-CGSTAB. 12 use of residual polynomials. 121 Breakdown. 118 nonlinear equations. 136 Bad Broyden method. 20 A-norm. 12 Approximate inverse. 126 Brsola algorithm. Index A-conjugacy.html. 37 Backtracking. 113 convergence linear equations. 6 Bi-CG. 123 sparse. 67 Contraction Mapping Theorem. 110 Bi-conjugacy. 123 restarting.com/SIAM/FR16. 48 algorithm. 137 Arnoldi process. 22 finite termination. 49 Chandrasekhar H-equation. 120 Broyden’s method. 34 Chord method. 3 Conjugate Gradient algorithm. 14 minimization property. 34 Contraction mapping. 25 CGNR. 25 CGS. 47 algorithm. 47 Bounded deterioration. 108 solution Broyden’s method. 48 algorithm.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 146 Newton-GMRES. 132 Banach Lemma. 123 limited memory. 6. 87 Newton-GMRES.

43 Globally convergent algorithm. 36 Induced matrix norm definition. 3 Inexact Newton method. 60 Kantorovich Theorem. 135. 80 nonlinearity. 40. 8 convergence result. 103 secant method. 54 Buy this book from SIAM at http://www. 135 GMRES algorithm. 37 INDEX loss of orthogonality. 104 Gauss–Seidel algorithm.ec-securehost. 39. 66 Forcing term. 87 solution Broyden’s method. 99 Newton’s method. 41 modified. 5 Iteration statistics. 55 GMRES. 39. 41 minimization property. 78 Hybrid algorithms. 110 Dennis–Mor´ Condition. 9 Jacobi preconditioning. 80 directional derivative. 40. 123 chord method. 77 contraction mapping. 9 Givens rotation. 100 Inverse power method. 136 Linear PDE Bi-CGSTAB. 95 choice of. 24 Jordan block.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 132 Inverse Tangent Newton–Armijo. 101 finite termination. 142 Line search. . 33 restarts. 34 GMRES(m). 94 Inverse secant equation. 46 loss of orthogonality. 46 algorithm. 40 H-equation. 107 H¨lder Continuity. 142 polynomial. 141 Newton-GMRES. 82 Shamanskii method.com/SIAM/FR16. 71 Newton–Armijo. 45 diagonalizable matrices. 143 two-point. 83 Krylov subspace. 114 e superlinear convergence. 127 CGNR. 164 nonlinear equations. 91 o High order methods. 123 Line Search parabolic three-point. 80 Elementary Jordan block. 96. This electronic version is for personal use and may not be duplicated or distributed.html. 57 Broyden’s method. 115 Diagonalizable matrix. 36 Limited memory. 79 DASPK. 95 Inner iteration. 33 Gram–Schmidt process. 80 Jacobian. 76. 34 Diagonalizing transformation. 128 Newton’s method. 67 finite difference Newton. 91 Jacobi algorithm. 60 Fixed point. 8 preconditioning. 11 Left preconditioning. 81 inexact Newton’s method. 66 Fixed-point iteration. 46 use of residual polynomials. 34 Difference approximation. 34 finite differences. 87 Newton-GMRES. 79 choice of h. 146 Iteration matrix.

11 Preconditioned CG Buy this book from SIAM at http://www. 58 Lipschitz continuous definition. 4 nonlinear as error estimate. 72 Residual polynomials application to CG. 72 Newton-GMRES. 81 stagnation.com/SIAM/FR16. 136 Newton’s method. This electronic version is for personal use and may not be duplicated or distributed. 136 Secant equation. 82 convergence. 24. 104 Picard iteration. 7 Modified Bratu Problem. 5 nonlinear. 71 finite differences. 14 application to CGNR. 113 Shamanskii method. 91 Secant update. 27. 36 Jacobi. 113 Quasi-residual. 27 for GMRES. 54 Polynomial preconditioning. 67 Local improvement. 86 Nsola algorithm. 36 Richardson iteration. 34 Nsol algorithm. 36.ec-securehost. 24. 19 GMRES. 91 r-order. 92 termination. 82 q-order. 101 convergence rates. 11 as error estimate. 51 Quasi-residual norm. 78 convergence. 127 TFQMR. 36 Richardson iteration. 103 finite differences. 79 solution Broyden’s method. 36.html. 9 Reorthogonalization. 100 Matrix norm. 6 Quasi-minimization. 36 incomplete factorization. 25 application to GMRES. 101 safeguarding. 74 convergence rate. 72 relative nonlinear. 33 application to TFQMR. 24. 23 Preconditioning conjugate gradient. 24 Poisson solver. 133 Search direction. 52 Quasi-Newton methods. 113 Secant method. 54 polynomial. 36 Schubert algorithm. 66 Poisson solver. 24. 78 algorithm. 7 Positive definite matrix. 132 Newton direction. 52 Relaxation parameter. 100 Over solving. 14 Richardson iteration. 105 Newton-iterative method. 24. 52 definition. 3 Matrix splitting. 66 Right preconditioning. 75 inexact. INDEX 165 algorithm. 81 inexact. 138 Nsolgm algorithm. 105 Outer iteration. . 71 algorithm. 27 preconditioner for CG. 41 Residual linear. 95 monotone convergence.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 100 Normal matrix.

This electronic version is for personal use and may not be duplicated or distributed. 7 Stagnation. 11 Symmetric positive definite matrix. 166 finite differences. 137 Symmetric Gauss–Seidel algorithm. 12 Spectral radius. 35. 53 finite difference. 5 as preconditioners. 4 nonlinear equations. 136 SOR algorithm. 12 Termination CG. 75 Stationary iterative methods. 66 Sufficient decrease. . 9 Symmetric matrix. 9 SPD matrix. 91 Sherman–Morrison formula. 16 GMRES. 38 linear equations.ec-securehost. 94 Newton’s method. 72 TFQMR.html. 9 preconditioning. 52 Weighted norm.Copyright ©1995 by the Society for Industrial and Applied Mathematics. 51 algorithm.com/SIAM/FR16. 124. 132 Simple decrease. 52 TFQMR. 98 INDEX Buy this book from SIAM at http://www. 132 Sherman–Morrison–Woodbury formula. 24 Successive substitution. 110 weights.

Sign up to vote on this title
UsefulNot useful