Kelley - Iterative Methods For Optimization-SIAM (1999) PDF

Copyright ©1999 by the Society for Industrial and Applied Mathematics.
This electronic version is for personal use and may not be duplicated or distributed.
Contents
Preface xiii
How to Get the Software xv
I Optimization of Smooth Functions 1

1 Basic Concepts 3
1.1 The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Necessary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Sufficient Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Quadratic Objective Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5.1 Positive Definite Hessian . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5.2 Indefinite Hessian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6.1 Discrete Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6.2 Parameter Identification . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6.3 Convex Quadratics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.7 Exercises on Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2 Local Convergence of Newton’s Method 13

2.1 Types of Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 The Standard Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Newton’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.1 Errors in Functions, Gradients, and Hessians . . . . . . . . . . . . . . 17
2.3.2 Termination of the Iteration . . . . . . . . . . . . . . . . . . . . . . . 21
2.4 Nonlinear Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 Gauss–Newton Iteration . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4.2 Overdetermined Problems . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.3 Underdetermined Problems . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Inexact Newton Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5.1 Convergence Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5.2 Implementation of Newton–CG . . . . . . . . . . . . . . . . . . . . . 30
2.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6.2 Discrete Control Problem . . . . . . . . . . . . . . . . . . . . . . . . 34
2.7 Exercises on Local Convergence . . . . . . . . . . . . . . . . . . . . . . . . . 35
ix
Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.
Copyright ©1999 by the Society for Industrial and Applied Mathematics. This electronic version is for personal use and may not be duplicated or distributed.
x CONTENTS
3 Global Convergence 39
3.1 The Method of Steepest Descent . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Line Search Methods and the Armijo Rule . . . . . . . . . . . . . . . . . . . . 40
3.2.1 Stepsize Control with Polynomial Models . . . . . . . . . . . . . . . . 43
3.2.2 Slow Convergence of Steepest Descent . . . . . . . . . . . . . . . . . 45
3.2.3 Damped Gauss–Newton Iteration . . . . . . . . . . . . . . . . . . . . 47
3.2.4 Nonlinear Conjugate Gradient Methods . . . . . . . . . . . . . . . . . 48
3.3 Trust Region Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3.1 Changing the Trust Region and the Step . . . . . . . . . . . . . . . . . 51
3.3.2 Global Convergence of Trust Region Algorithms . . . . . . . . . . . . 52
3.3.3 A Unidirectional Trust Region Algorithm . . . . . . . . . . . . . . . . 54
3.3.4 The Exact Solution of the Trust Region Problem . . . . . . . . . . . . 55
3.3.5 The Levenberg–Marquardt Parameter . . . . . . . . . . . . . . . . . . 56
3.3.6 Superlinear Convergence: The Dogleg . . . . . . . . . . . . . . . . . . 58
3.3.7 A Trust Region Method for Newton–CG . . . . . . . . . . . . . . . . . 63
3.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 Exercises on Global Convergence . . . . . . . . . . . . . . . . . . . . . . . . 68
4 The BFGS Method 71

4.1 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.1.1 Local Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.1.2 Global Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.2.1 Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.2.2 A BFGS–Armijo Algorithm . . . . . . . . . . . . . . . . . . . . . . . 80
4.3 Other Quasi-Newton Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4.1 Parameter ID Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.5 Exercises on BFGS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5 Simple Bound Constraints 87

5.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2 Necessary Conditions for Optimality . . . . . . . . . . . . . . . . . . . . . . . 87
5.3 Sufficient Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.4 The Gradient Projection Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 91
5.4.1 Termination of the Iteration . . . . . . . . . . . . . . . . . . . . . . . 91
5.4.2 Convergence Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4.3 Identification of the Active Set . . . . . . . . . . . . . . . . . . . . . . 95
5.4.4 A Proof of Theorem 5.2.4 . . . . . . . . . . . . . . . . . . . . . . . . 96
5.5 Superlinear Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.5.1 The Scaled Gradient Projection Algorithm . . . . . . . . . . . . . . . . 96
5.5.2 The Projected Newton Method . . . . . . . . . . . . . . . . . . . . . . 100
5.5.3 A Projected BFGS–Armijo Algorithm . . . . . . . . . . . . . . . . . . 102
5.6 Other Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.6.1 Infinite-Dimensional Problems . . . . . . . . . . . . . . . . . . . . . . 106
5.7 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.7.1 Parameter ID Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 106

CONTENTS xi
5.8 Exercises on Bound Constrained Optimization . . . . . . . . . . . . . . . . . . 108
II Optimization of Noisy Functions 109

6 Basic Concepts and Goals 111
6.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.2 The Simplex Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.2.1 Forward Difference Simplex Gradient . . . . . . . . . . . . . . . . . . 113
6.2.2 Centered Difference Simplex Gradient . . . . . . . . . . . . . . . . . . 115
6.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.3.1 Weber’s Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.3.2 Perturbed Convex Quadratics . . . . . . . . . . . . . . . . . . . . . . 119
6.3.3 Lennard–Jones Problem . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.4 Exercises on Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7 Implicit Filtering 123

7.1 Description and Analysis of Implicit Filtering . . . . . . . . . . . . . . . . . . 123
7.2 Quasi-Newton Methods and Implicit Filtering . . . . . . . . . . . . . . . . . . 124
7.3 Implementation Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.4 Implicit Filtering for Bound Constrained Problems . . . . . . . . . . . . . . . 126
7.5 Restarting and Minima at All Scales . . . . . . . . . . . . . . . . . . . . . . . 127
7.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.6.1 Weber’s Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.6.2 Parameter ID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.7 Exercises on Implicit Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8 Direct Search Algorithms 135

8.1 The Nelder–Mead Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.1.1 Description and Implementation . . . . . . . . . . . . . . . . . . . . . 135
8.1.2 Sufficient Decrease and the Simplex Gradient . . . . . . . . . . . . . . 137
8.1.3 McKinnon’s Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.1.4 Restarting the Nelder–Mead Algorithm . . . . . . . . . . . . . . . . . 141
8.2 Multidirectional Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.2.2 Convergence and the Simplex Gradient . . . . . . . . . . . . . . . . . 144
8.3 The Hooke–Jeeves Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
8.3.2 Convergence and the Simplex Gradient . . . . . . . . . . . . . . . . . 148
8.4 Other Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.4.1 Surrogate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.4.2 The DIRECT Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.5.1 Weber’s Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.5.2 Parameter ID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
8.6 Exercises on Search Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 159

xii CONTENTS
Bibliography 161
Index 177

Preface
This book on unconstrained and bound constrained optimization can be used as a tutorial for
self-study or a reference by those who solve such problems in their work. It can also serve as a
textbook in an introductory optimization course.
As in my earlier book [154] on linear and nonlinear equations, we treat a small number of
methods in depth, giving a less detailed description of only a few (for example, the nonlinear
conjugate gradient method and the DIRECT algorithm). We aim for clarity and brevity rather
than complete generality and confine our scope to algorithms that are easy to implement (by the
reader!) and understand.
One consequence of this approach is that the algorithms in this book are often special cases
of more general ones in the literature. For example, in Chapter 3, we provide details only
for trust region globalizations of Newton’s method for unconstrained problems and line search
globalizations of the BFGS quasi-Newton method for unconstrained and bound constrained
problems. We refer the reader to the literature for more general results. Our intention is that
both our algorithms and proofs, being special cases, are more concise and simple than others in
the literature and illustrate the central issues more clearly than a fully general formulation.
Part II of this book covers some algorithms for noisy or global optimization or both. There
are many interesting algorithms in this class, and this book is limited to those deterministic
algorithms that can be implemented in a more-or-less straightforward way. We do not, for
example, cover simulated annealing, genetic algorithms, response surface methods, or random
search procedures.
The reader of this book should be familiar with the material in an elementary graduate level
course in numerical analysis, in particular direct and iterative methods for the solution of linear
equations and linear least squares problems. The material in texts such as [127] and [264] is
sufficient.
A suite of MATLAB∗ codes has been written to accompany this book. These codes were
used to generate the computational examples in the book, but the algorithms do not depend
on the MATLAB environment and the reader can easily implement the algorithms in another
language, either directly from the algorithmic descriptions or by translating the MATLAB code.
The MATLAB environment is an excellent choice for experimentation, doing the exercises, and
small-to-medium-scale production work. Large-scale work on high-performance computers is
best done in another language. The reader should also be aware that there is a large amount of
high-quality software available for optimization. The book [195], for example, provides pointers
to several useful packages.
Parts of this book are based upon work supported by the National Science Foundation over
several years, most recently under National Science Foundation grants DMS-9321938, DMS-
9700569, and DMS-9714811, and by allocations of computing resources from the North Carolina
Supercomputing Center. Any opinions, findings, and conclusions or recommendations expressed
∗ MATLAB is a registered trademark of The MathWorks, Inc., 24 Prime Park Way, Natick, MA 01760, USA, (508)
653-1415, [email protected], http://www.mathworks.com.
xiii
xiv PREFACE
in this material are those of the author and do not necessarily reflect the views of the National
Science Foundation or of the North Carolina Supercomputing Center.
The list of students and colleagues who have helped me with this project, directly, through
collaborations/discussions on issues that I treat in the manuscript, by providing pointers to the
literature, or as a source of inspiration, is long. I am particularly indebted to Tom Banks, Jim
Banoczi, John Betts, David Bortz, Steve Campbell, Tony Choi, Andy Conn, Douglas Cooper, Joe
David, John Dennis, Owen Eslinger, Jörg Gablonsky, Paul Gilmore, Matthias Heinkenschloß,
Laura Helfrich, Lea Jenkins, Vickie Kearn, Carl and Betty Kelley, Debbie Lockhart, Casey Miller,
Jorge Moré, Mary Rose Muccie, John Nelder, Chung-Wei Ng, Deborah Poulson, Ekkehard
Sachs, Dave Shanno, Joseph Skudlarek, Dan Sorensen, John Strikwerda, Mike Tocci, Jon Tolle,
Virginia Torczon, Floria Tosca, Hien Tran, Margaret Wright, Steve Wright, and Kevin Yoemans.
C. T. Kelley
Raleigh, North Carolina

How to Get the Software
All computations reported in this book were done in MATLAB (version 5.2 on various SUN
SPARCstations and on an Apple Macintosh Powerbook 2400). The suite of MATLAB codes that
we used for the examples is available by anonymous ftp from ftp.math.ncsu.edu in the directory
FTP/kelley/optimization/matlab
or from SIAM’s World Wide Web server at
http://www.siam.org/books/fr18/
One can obtain MATLAB from

The MathWorks, Inc.
3 Apple Hill Drive
Natick, MA 01760-2098
(508) 647-7000
Fax: (508) 647-7001
E-mail: [email protected]
WWW: http://www.mathworks.com
xv
Part I
Optimization of Smooth Functions


Chapter 1
Basic Concepts
1.1 The Problem

The unconstrained optimization problem is to minimize a real-valued function f of N variables.
By this we mean to find a local minimizer, that is, a point x∗ such that
(1.1) f (x∗ ) ≤ f (x) for all x near x∗ .
It is standard to express this problem as
(1.2) min f (x)

x
or to say that we seek to solve the problem min f . The understanding is that (1.1) means that we
seek a local minimizer. We will refer to f as the objective function and to f (x∗ ) as the minimum
or minimum value. If a local minimizer x∗ exists, we say a minimum is attained at x∗ .
We say that problem (1.2) is unconstrained because we impose no conditions on the inde-
pendent variables x and assume that f is defined for all x.
The local minimization problem is different from (and much easier than) the global mini-
mization problem in which a global minimizer, a point x∗ such that
(1.3) f (x∗ ) ≤ f (x) for all x,
is sought.
The constrained optimization problem is to minimize a function f over a set U ⊂ RN . A
local minimizer, therefore, is an x∗ ∈ U such that
(1.4) f (x∗ ) ≤ f (x) for all x ∈ U near x∗ .
Similar to (1.2) we express this as

(1.5) min f (x)
x∈U
or say that we seek to solve the problem minU f . A global minimizer is a point x∗ ∈ U such
that
(1.6) f (x∗ ) ≤ f (x) for all x ∈ U .
We consider only the simplest constrained problems in this book (Chapter 5 and §7.4) and refer
the reader to [104], [117], [195], and [66] for deeper discussions of constrained optimization
and pointers to software.
Having posed an optimization problem one can proceed in the classical way and use methods
that require smoothness of f . That is the approach we take in this first part of the book. These
3
4 ITERATIVE METHODS FOR OPTIMIZATION
methods can fail if the objective function has discontinuities or irregularities. Such nonsmooth
effects are common and can be caused, for example, by truncation error in internal calculations
for f , noise in internal probabilistic modeling in f , table lookup within f , or use of experimental
data in f . We address a class of methods for dealing with such problems in Part II.
1.2 Notation
In this book, following the convention in [154], vectors are to be understood as column vectors.
The vector x∗ will denote a solution, x a potential solution, and {xk }k≥0 the sequence of iterates.
We will refer to x0 as the initial iterate. x0 is sometimes timidly called the initial guess. 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 . We will rarely need to refer to individual components of vectors. We will let
∂f /∂xi denote the partial derivative of f with respect to (x)i . As is standard [154], e = x − x∗
will denote the error, en = xn − x∗ the error in the nth iterate, and B(r) the ball of radius r
about x∗
B(r) = {x |
e
< r}.
For x ∈ RN we let ∇f (x) ∈ RN denote the gradient of f ,
∇f (x) = (∂f /∂x1 , . . . , ∂f /∂xN ),
when it exists.
We let ∇2 f denote the Hessian of f ,
(∇2 f )ij = ∂ 2 f /∂xi ∂xj ,
when it exists. Note that ∇2 f is the Jacobian of ∇f . However, ∇2 f has more structure than
a Jacobian for a general nonlinear function. If f is twice continuously differentiable, then the
Hessian is symmetric ((∇2 f )ij = (∇2 f )ji ) by equality of mixed partial derivatives [229].
In this book we will consistently use the Euclidean norm

N

x
= (x)2i .
i=1
When we refer to a matrix norm we will mean the matrix norm induced by the Euclidean norm

Ax

A
= max .
x=0
x
In optimization definiteness or semidefiniteness of the Hessian plays an important role in

the necessary and sufficient conditions for optimality that we discuss in §1.3 and 1.4 and in our
choice of algorithms throughout this book.
Definition 1.2.1. An N × N matrix A is positive semidefinite if xT Ax ≥ 0 for all x ∈ RN .
A is positive definite if xT Ax > 0 for all x ∈ RN , x = 0. If A has both positive and negative
eigenvalues, we say A is indefinite. If A is symmetric and positive definite, we will say A is spd.
We will use two forms of the fundamental theorem of calculus, one for the function–gradient
pair and one for the gradient–Hessian.
Theorem 1.2.1. Let f be twice continuously differentiable in a neighborhood of a line
segment between points x∗ , x = x∗ + e ∈ RN ; then
1
f (x) = f (x∗ ) + ∇f (x∗ + te)T e dt
0

BASIC CONCEPTS 5
and 1
∇f (x) = ∇f (x∗ ) + ∇2 f (x∗ + te)e dt.
0
A direct consequence (see Exercise 1.7.1) of Theorem 1.2.1 is the following form of Taylor’s
theorem we will use throughout this book.
Theorem 1.2.2. Let f be twice continuously differentiable in a neighborhood of a point
x∗ ∈ RN . Then for e ∈ RN and
e
sufficiently small
(1.7) f (x∗ + e) = f (x∗ ) + ∇f (x∗ )T e + eT ∇2 f (x∗ )e/2 + o(

e
2 ).
1.3 Necessary Conditions

Let f be twice continuously differentiable. We will use Taylor’s theorem in a simple way to
show that the gradient of f vanishes at a local minimizer and the Hessian is positive semidefinite.
These are the necessary conditions for optimality.
The necessary conditions relate (1.1) to a nonlinear equation and allow one to use fast al-
gorithms for nonlinear equations [84], [154], [211] to compute minimizers. Therefore, the
necessary conditions for optimality will be used in a critical way in the discussion of local con-
vergence in Chapter 2. A critical first step in the design of an algorithm for a new optimization
problem is the formulation of necessary conditions. Of course, the gradient vanishes at a maxi-
mum, too, and the utility of the nonlinear equations formulation is restricted to a neighborhood
of a minimizer.
Theorem 1.3.1. Let f be twice continuously differentiable and let x∗ be a local minimizer
of f . Then
∇f (x∗ ) = 0.
Moreover ∇2 f (x∗ ) is positive semidefinite.
Proof. Let u ∈ RN be given. Taylor’s theorem states that for all real t sufficiently small
t2 T 2
f (x∗ + tu) = f (x∗ ) + t∇f (x∗ )T u + u ∇ f (x∗ )u + o(t2 ).
2
Since x∗ is a local minimizer we must have for t sufficiently small 0 ≤ f (x∗ + tu) − f (x∗ ) and
hence
t
(1.8) ∇f (x∗ )T u + uT ∇2 f (x∗ )u + o(t) ≥ 0
2
for all t sufficiently small and all u ∈ RN . So if we set t = 0 and u = −∇f (x∗ ) we obtain

∇f (x∗ )
2 = 0.
Setting ∇f (x∗ ) = 0, dividing by t, and setting t = 0 in (1.8), we obtain

1 T 2
u ∇ f (x∗ )u ≥ 0
2
for all u ∈ RN . This completes the proof.
The condition that ∇f (x∗ ) = 0 is called the first-order necessary condition and a point
satisfying that condition is called a stationary point or a critical point.

1.4 Sufficient Conditions

A stationary point need not be a minimizer. For example, the function φ(t) = −t4 satisfies the
necessary conditions at 0, which is a maximizer of φ. To obtain a minimizer we must require that
the second derivative be nonnegative. This alone is not sufficient (think of φ(t) = t3 ) and only
if the second derivative is strictly positive can we be completely certain. These are the sufficient
conditions for optimality.
Theorem 1.4.1. Let f be twice continuously differentiable in a neighborhood of x∗ . Assume

that ∇f (x∗ ) = 0 and that ∇2 f (x∗ ) is positive definite. Then x∗ is a local minimizer of f .
Proof. Let 0 = u ∈ RN . For sufficiently small t we have
t2 T 2
f (x∗ + tu) = f (x∗ ) + t∇f (x∗ )T u + u ∇ f (x∗ )u + o(t2 )
2
t2 T 2
= f (x∗ ) + u ∇ f (x∗ )u + o(t2 ).
2
Hence, if λ > 0 is the smallest eigenvalue of ∇2 f (x∗ ) we have
λ
f (x∗ + tu) − f (x∗ ) ≥
tu
2 + o(t2 ) > 0
2
for t sufficiently small. Hence x∗ is a local minimizer.
1.5 Quadratic Objective Functions

The simplest optimization problems are those with quadratic objective functions. Here
1
(1.9) f (x) = −xT b + xT Hx.
2
We may, without loss of generality, assume that H is symmetric because

T T H + HT
(1.10) x Hx = x x.
2
Quadratic functions form the basis for most of the algorithms in Part I, which approximate an
objective function f by a quadratic model and minimize that model. In this section we discuss
some elementary issues in quadratic optimization.
Clearly,
∇2 f (x) = H
for all x. The symmetry of H implies that
∇f (x) = −b + Hx.
Definition 1.5.1. The quadratic function f in (1.9) is convex if H is positive semidefinite.

BASIC CONCEPTS 7
1.5.1 Positive Definite Hessian

The necessary conditions for optimality imply that if a quadratic function f has a local minimum
x∗ , then H is positive semidefinite and
(1.11) Hx∗ = b.
In particular, if H is spd (and hence nonsingular), the unique global minimizer is the solution of
the linear system (1.11).
If H is a dense matrix and N is not too large, it is reasonable to solve (1.11) by computing
the Cholesky factorization [249], [127] of H
H = LLT ,
where L is a nonsingular lower triangular matrix with positive diagonal, and then solving (1.11)
by two triangular solves. If H is indefinite the Cholesky factorization will not exist and the
standard implementation [127], [249], [264] will fail because the computation of the diagonal
of L will require a real square root of a negative number or a division by zero.
If N is very large, H is sparse, or a matrix representation of H is not available, a more
efficient approach is the conjugate gradient iteration [154], [141]. This iteration requires only
matrix–vector products, a feature which we will use in a direct way in §§2.5 and 3.3.7. Our
formulation of the algorithm uses x as both an input and output variable. On input x contains
x0 , the initial iterate, and on output the approximate solution. We terminate the iteration if the
relative residual is sufficiently small, i.e.,

b − Hx
≤
b
or if too many iterations have been taken.

Algorithm 1.5.1. cg(x, b, H, , kmax)
1. r = b − Hx, ρ0 =
r
2 , k = 1.
√
2. Do While ρk−1 >
b
and k < kmax
(a) if k = 1 then p = r
else
β = ρk−1 /ρk−2 and p = r + βp
(b) w = Hp
(c) α = ρk−1 /pT w
(d) x = x + αp
(e) r = r − αw
(f) ρk =
r
2
(g) k = k + 1
Note that if H is not spd, the denominator in α = ρk−1 /pT w may vanish, resulting in
breakdown of the iteration.
The conjugate gradient iteration minimizes f over an increasing sequence of nested subspaces
of RN [127], [154]. We have that
f (xk ) ≤ f (x) for all x ∈ x0 + Kk ,

where Kk is the Krylov subspace
Kk = span(r0 , Hr0 , . . . , H k−1 r0 )
for k ≥ 1.
While in principle the iteration must converge after N iterations and conjugate gradient can
be regarded as a direct solver, N is, in practice, far too many iterations for the large problems to
which conjugate gradient is applied. As an iterative method, the performance of the conjugate
gradient algorithm depends both on b and on the spectrum of H (see [154] and the references
cited therein). A general convergence estimate [68], [60], which will suffice for the discussion
here, is

k
∗ ∗ κ(H) − 1
(1.12)
xk − x
H ≤ 2
x0 − x
H .
κ(H) + 1
In (1.12), the H-norm of a vector is defined by

u
2H = uT Hu
for an spd matrix H. κ(H) is the l2 condition number
κ(H) =
H
H −1
.
For spd H
κ(H) = λl λ−1
s ,
where λl and λs are the largest and smallest eigenvalues of H. Geometrically, κ(H) is large if
the ellipsoidal level surfaces of f are very far from spherical.
The conjugate gradient iteration will perform well if κ(H) is near 1 and may perform very
poorly if κ(H) is large. The performance can be improved by preconditioning, which transforms
(1.11) into one with a coefficient matrix having eigenvalues near 1. Suppose that M is spd and
a sufficiently good approximation to H −1 so that
κ(M 1/2 HM 1/2 )
is much smaller that κ(H). In that case, (1.12) would indicate that far fewer conjugate gradient
iterations might be needed to solve
(1.13) M 1/2 HM 1/2 z = M 1/2 b
than to solve (1.11). Moreover, the solution x∗ of (1.11) could be recovered from the solution
z ∗ of (1.13) by
(1.14) x = M 1/2 z.
In practice, the square root of the preconditioning matrix M need not be computed. The algo-
rithm, using the same conventions that we used for cg, is
Algorithm 1.5.2. pcg(x, b, H, M, , kmax)
1. r = b − Hx, ρ0 =
r
2 , k = 1
√
2. Do While ρk−1 >
b
and k < kmax
(a) z = M r
(b) τk−1 = z T r

BASIC CONCEPTS 9
(c) if k = 1 then β = 0 and p = z

else
β = τk−1 /τk−2 , p = z + βp
(d) w = Hp
(e) α = τk−1 /pT w
(f) x = x + αp
(g) r = r − αw
(h) ρk = rT r
(i) k = k + 1
Note that only products of M with vectors in RN are needed and that a matrix representation
of M need not be stored. We refer the reader to [11], [15], [127], and [154] for more discussion
of preconditioners and their construction.
1.5.2 Indefinite Hessian

If H is indefinite, the necessary conditions, Theorem 1.3.1, imply that there will be no local
minimum. Even so, it will be important to understand some properties of quadratic problems
with indefinite Hessians when we design algorithms with initial iterates far from local minimizers
and we discuss some of the issues here.
If
uT Hu < 0,
we say that u is a direction of negative curvature. If u is a direction of negative curvature, then
f (x + tu) will decrease to −∞ as t → ∞.
1.6 Examples
It will be useful to have some example problems to solve as we develop the algorithms. The
examples here are included to encourage the reader to experiment with the algorithms and play
with the MATLAB codes. The codes for the problems themselves are included with the set of
MATLAB codes. The author of this book does not encourage the reader to regard the examples
as anything more than examples. In particular, they are not real-world problems, and should not
be used as an exhaustive test suite for a code. While there are documented collections of test
problems (for example, [10] and [26]), the reader should always evaluate and compare algorithms
in the context of his/her own problems.
Some of the problems are directly related to applications. When that is the case we will cite
some of the relevant literature. Other examples are included because they are small, simple, and
illustrate important effects that can be hidden by the complexity of more serious problems.
1.6.1 Discrete Optimal Control

This is a classic example of a problem in which gradient evaluations cost little more than function
evaluations.
We begin with the continuous optimal control problems and discuss how gradients are com-
puted and then move to the discretizations. We will not dwell on the functional analytic issues
surrounding the rigorous definition of gradients of maps on function spaces, but the reader should
be aware that control problems require careful attention to this. The most important results can

be found in [151]. The function space setting for the particular control problems of interest in this
section can be found in [170], [158], and [159], as can a discussion of more general problems.
The infinite-dimensional problem is
(1.15) min f,
u
where T
(1.16) f (u) = L(y(t), u(t), t) dt,
0
and we seek an optimal point u ∈ L∞ [0, T ]. u is called the control variable or simply the
control. The function L is given and y, the state variable, satisfies the initial value problem
(with ẏ = dy/dt)
(1.17) ẏ(t) = φ(y(t), u(t), t), y(0) = y0 .
One could view the problem (1.15)–(1.17) as a constrained optimization problem or, as we
do here, think of the evaluation of f as requiring the solution of (1.17) before the integral on the
right side of (1.16) can be evaluated. This means that evaluation of f requires the solution of
(1.17), which is called the state equation.
∇f (u), the gradient of f at u with respect to the L2 inner product, is uniquely determined,
if it exists, by
T
(1.18) f (u + w) − f (u) − (∇f (u))(t)w(t) dt = o(
w
)
0
as
w
→ 0, uniformly in w. If ∇f (u) exists then

T
df (u + ξw)
(∇f (u))(t)w(t) dt = .
0 dξ ξ=0
If L and φ are continuously differentiable, then ∇f (u), as a function of t, is given by
(1.19) ∇f (u)(t) = p(t)φu (y(t), u(t), t) + Lu (y(t), u(t), t).
In (1.19) p, the adjoint variable, satisfies the final-value problem on [0, T ]
(1.20) −ṗ(t) = p(t)φy (y(t), u(t), t) + Ly (y(t), u(t), t), p(T ) = 0.
So computing the gradient requires u and y, hence a solution of the state equation, and p, which
requires a solution of (1.20), a final-value problem for the adjoint equation. In the general case,
(1.17) is nonlinear, but (1.20) is a linear problem for p, which should be expected to be easier
to solve. This is the motivation for our claim that a gradient evaluation costs little more than a
function evaluation.
The discrete problems of interest here are constructed by solving (1.17) by numerical in-
tegration. After doing that, one can derive an adjoint variable and compute gradients using a
discrete form of (1.19). However, in [139] the equation for the adjoint variable of the discrete
problem is usually not a discretization of (1.20). For the forward Euler method, however, the
discretization of the adjoint equation is the adjoint equation for the discrete problem and we use
that discretization here for that reason.
The fully discrete problem is minu f , where u ∈ RN and
N

f (u) = L((y)j , (u)j , j),
j=1

BASIC CONCEPTS 11
and the states {xj } are given by the Euler recursion
yj+1 = yj + hφ((y)j , (u)j , j) for j = 0, . . . , N − 1,
where h = T /(N − 1) and x0 is given. Then
(∇f (u))j = (p)j φu ((y)j , (u)j , j) + Lu ((y)j , (u)j , j),
where (p)N = 0 and

(p)j−1 = (p)j + h (p)j φy ((y)j , (u)j , j) + Ly ((y)j , (u)j , j) for j = N, . . . , 1.
1.6.2 Parameter Identification

This example, taken from [13], will appear throughout the book. The problem is small with
N = 2. The goal is to identify the damping c and spring constant k of a linear spring by
minimizing the difference of a numerical prediction and measured data. The experimental
scenario is that the spring-mass system will be set into motion by an initial displacement from
equilibrium and measurements of displacements will be taken at equally spaced increments in
time.
The motion of an unforced harmonic oscillator satisfies the initial value problem
(1.21) u + cu + ku = 0; u(0) = u0 , u (0) = 0,
on the interval [0, T ]. We let x = (c, k)T be the vector of unknown parameters and, when the
dependence on the parameters needs to be explicit, we will write u(t : x) instead of u(t) for the
solution of (1.21). If the displacement is sampled at {tj }M j=1 , where tj = (j − 1)T /(M − 1),
and the observations for u are {uj }Mj=1 , then the objective function is
M
1
(1.22) f (x) = |u(tj : x) − uj |2 .
2 j=1
This is an example of a nonlinear least squares problem.

u is differentiable with respect to x when c2 − 4k = 0. In that case, the gradient of f is

M ∂u(tj :x)
j=1 ∂c (u(tj : x) − uj )
(1.23) ∇f (x) = M ∂u(tj :x) .
j=1 ∂k (u(tj : x) − uj )
We can compute the derivatives of u(t : x) with respect to the parameters by solving the
sensitivity equations. Differentiating (1.21) with respect to c and k and setting w1 = ∂u/∂c and
w2 = ∂u/∂k we obtain
w1 + u + cw1 + kw1 = 0; w1 (0) = w1 (0) = 0,

(1.24)
w2 + cw2 + u + kw2 = 0; w2 (0) = w2 (0) = 0.
If c is large, the initial value problems (1.21) and (1.24) will be stiff and one should use a good
variable step stiff integrator. We refer the reader to [110], [8], [235] for details on this issue. In
the numerical examples in this book we used the MATLAB code ode15s from [236]. Stiffness
can also arise in the optimal control problem from §1.6.1 but does not in the specific examples
we use in this book. We caution the reader that when one uses an ODE code the results may only
be expected to be accurate to the tolerances input to the code. This limitation on the accuracy
must be taken into account, for example, when approximating the Hessian by differences.

1.6.3 Convex Quadratics

While convex quadratic problems are, in a sense, the easiest of optimization problems, they
present surprising challenges to the sampling algorithms presented in Part II and can illustrate
fundamental problems with classical gradient-based methods like the steepest descent algorithm
from §3.1. Our examples will all take N = 2, b = 0, and

λs 0
H= ,
0 λl
where 0 < λs ≤ λl . The function to be minimized is
f (x) = xT Hx
and the minimizer is x∗ = (0, 0)T .

As λl /λs becomes large, the level curves of f become elongated. When λs = λl = 1,
minx f is the easiest problem in optimization.
1.7 Exercises on Basic Concepts

1.7.1. Prove Theorem 1.2.2.
1.7.2. Consider the parameter identification problem for x = (c, k, ω, φ)T ∈ R4 associated with
the initial value problem
u + cu + ku = sin(ωt + φ); u(0) = 10, u (0) = 0.
For what values of x is u differentiable? Derive the sensitivity equations for those values
of x for which u is differentiable.
1.7.3. Solve the system of sensitivity equations from exercise 1.7.2 numerically for c = 10,
k = 1, ω = π, and φ = 0 using the integrator of your choice. What happens if you use a
nonstiff integrator?
1.7.4. Let N = 2, d = (1, 1)T , and let f (x) = xT d + xT x. Compute, by hand, the minimizer
using conjugate gradient iteration.
1.7.5. For the same f as in exercise 1.7.4 solve the constrained optimization problem
min f (x),
x∈U
where U is the circle centered at (0, 0)T of radius 1/3. You can solve this by inspection;
no computer and very little mathematics is needed.

Chapter 2
Local Convergence of Newton’s

Method
By a local convergence method we mean one that requires that the initial iterate x0 is close to a
local minimizer x∗ at which the sufficient conditions hold.
2.1 Types of Convergence

We begin with the standard taxonomy of convergence rates [84], [154], [211].
Definition 2.1.1. 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

xn+1 − x∗
lim = 0.
n→∞
xn − x∗
• xn → x∗ q-linearly with q-factor σ ∈ (0, 1) if

xn+1 − x∗
≤ σ
xn − x∗
for n sufficiently large.
Definition 2.1.2. 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.
We remind the reader 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 of 2.
13
In some cases the accuracy of the iteration can be improved by means that are external
to the algorithm, say, by evaluation of the objective function and its gradient with increasing
accuracy as the iteration progresses. In such cases, one has no guarantee that the accuracy of
the iteration is monotonically increasing but only that the accuracy of the results is improving at
a rate determined by the improving accuracy in the function–gradient evaluations. The concept
of r-type convergence captures this effect.
Definition 2.1.3. Let {xn } ⊂ RN and x∗ ∈ RN . Then {xn } converges to x∗ r-( quadrat-
ically, superlinearly, linearly) if there is a sequence {ξn } ⊂ R converging q-(quadratically,
superlinearly, linearly) to 0 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.
2.2 The Standard Assumptions

We will assume that local minimizers satisfy the standard assumptions which, like the standard
assumptions for nonlinear equations in [154], will guarantee that Newton’s method converges
q-quadratically to x∗ . We will assume throughout this book that f and x∗ satisfy Assumption
2.2.1.
Assumption 2.2.1.
1. f is twice differentiable and
(2.1)
∇2 f (x) − ∇2 f (y)
≤ γ
x − y
.
2. ∇f (x∗ ) = 0.
3. ∇2 f (x∗ ) is positive definite.
We sometimes say that f is twice Lipschitz continuously differentiable with Lipschitz constant
γ to mean that part 1 of the standard assumptions holds.
If the standard assumptions hold then Theorem 1.4.1 implies that x∗ is a local minimizer
of f . Moreover, the standard assumptions for nonlinear equations [154] hold for the system
∇f (x) = 0. This means that all of the local convergence results for nonlinear equations can be
applied to unconstrained optimization problems. In this chapter we will quote those results from
nonlinear equations as they apply to unconstrained optimization. However, these statements
must be understood in the context of optimization. We will use, for example, the fact that the
Hessian (the Jacobian of ∇f ) is positive definite at x∗ in our solution of the linear equation for
the Newton step. We will also use this in our interpretation of the Newton iteration.
2.3 Newton’s Method

As in [154] we will define iterative methods in terms of the transition from a current iteration xc
to a new one x+ . In the case of a system of nonlinear equations, for example, x+ is the root of
the local linear model of F about xc
Mc (x) = F (xc ) + F (xc )(x − xc ).

LOCAL CONVERGENCE 15
Solving Mc (x+ ) = 0 leads to the standard formula for the Newton iteration
(2.2) x+ = xc − F (xc )−1 F (xc ).
One could say that Newton’s method for unconstrained optimization is simply the method
for nonlinear equations applied to ∇f (x) = 0. While this is technically correct if xc is near a
minimizer, it is utterly wrong if xc is near a maximum. A more precise way of expressing the
idea is to say that x+ is a minimizer of the local quadratic model of f about xc .
1
mc (x) = f (xc ) + ∇f (xc )T (x − xc ) + (x − xc )T ∇2 f (xc )(x − xc ).
2
If ∇2 f (xc ) is positive definite, then the minimizer x+ of mc is the unique solution of ∇mc (x) =
0. Hence,
0 = ∇mc (x+ ) = ∇f (xc ) + ∇2 f (xc )(x+ − xc ).
Therefore,
(2.3) x+ = xc − (∇2 f (xc ))−1 ∇f (xc ),
which is the same as (2.2) with F replaced by ∇f and F by ∇2 f . Of course, x+ is not computed
by forming an inverse matrix. Rather, given xc , ∇f (xc ) is computed and the linear equation
(2.4) ∇2 f (xc )s = −∇f (xc )
is solved for the step s. Then (2.3) simply says that x+ = xc + s.

However, if uc is far from a minimizer, ∇2 f (uc ) could have negative eigenvalues and the
quadratic model will not have local minimizers (see exercise 2.7.4), and Mc , the local linear
model of ∇f about uc , could have roots which correspond to local maxima or saddle points
of mc . Hence, we must take care when far from a minimizer in making a correspondence
between Newton’s method for minimization and Newton’s method for nonlinear equations. In
this chapter, however, we will assume that we are sufficiently near a local minimizer for the
standard assumptions for local optimality to imply those for nonlinear equations (as applied to
∇f ). Most of the proofs in this chapter are very similar to the corresponding results, [154], for
nonlinear equations. We include them in the interest of completeness.
We begin with a lemma from [154], which we state without proof.
Lemma 2.3.1. Assume that the standard assumptions hold. Then there is δ > 0 so that for
all x ∈ B(δ)
(2.5)
∇2 f (x)
≤ 2
∇2 f (x∗ )
,
(2.6)
(∇2 f (x))−1
≤ 2
(∇2 f (x∗ ))−1
,
and
(2.7)
(∇2 f (x∗ ))−1
−1
e
/2 ≤
∇f (x)
≤ 2
∇2 f (x∗ )
e
.
As a first example, we prove the local convergence for Newton’s method.

Theorem 2.3.2. Let the standard assumptions hold. Then there are K > 0 and δ > 0 such
that if xc ∈ B(δ), the Newton iterate from xc given by (2.3) satisfies
(2.8)
e+
≤ K
ec
2 .
Proof. Let δ be small enough so that the conclusions of Lemma 2.3.1 hold. By Theorem 1.2.1
1
e+ = ec − ∇2 f (xc )−1 ∇f (xc ) = (∇2 f (xc ))−1 (∇2 f (xc ) − ∇2 f (x∗ + tec ))ec dt.
0

By Lemma 2.3.1 and the Lipschitz continuity of ∇2 f ,

e+
≤ (2
(∇2 f (x∗ ))−1
)γ
ec
2 /2.
This completes the proof of (2.8) with K = γ

(∇2 f (x∗ ))−1
.
As in the nonlinear equations setting, Theorem 2.3.2 implies that the complete iteration is
locally quadratically convergent.
Theorem 2.3.3. Let the standard assumptions hold. Then there is δ > 0 such that if
x0 ∈ B(δ), the Newton iteration
xn+1 = xn − (∇2 f (xn ))−1 ∇f (xn )
converges q-quadratically to x∗ .
Proof. Let δ be small enough so that the conclusions of Theorem 2.3.2 hold. Reduce δ if
needed so that Kδ = η < 1. Then if n ≥ 0 and xn ∈ B(δ), Theorem 2.3.2 implies that
(2.9)
en+1
≤ K
en
2 ≤ η
en
<
en
and hence xn+1 ∈ B(ηδ) ⊂ B(δ). Therefore, if xn ∈ B(δ) we may continue the iteration. Since
x0 ∈ B(δ) by assumption, the entire sequence {xn } ⊂ B(δ). (2.9) then implies that xn → x∗
q-quadratically.
Newton’s method, from the local convergence point of view, is exactly the same as that
for nonlinear equations applied to the problem of finding a root of ∇f . We exploit the extra
structure of positive definiteness of ∇2 f with an implementation of Newton’s method based on
the Cholesky factorization [127], [249], [264]
(2.10) ∇2 f (u) = LLT ,
where L is lower triangular and has a positive diagonal.

We terminate the iteration when ∇f is sufficiently small (see [154]). A natural criterion is
to demand a relative decrease in
∇f
and terminate when
(2.11)
∇f (xn )
≤ τr
∇f (x0 )
,
where τr ∈ (0, 1) is the desired reduction in the gradient norm. However, if

∇f (x0 )
is very
small, it may not be possible to satisfy (2.11) in floating point arithmetic and an algorithm based
entirely on (2.11) might never terminate. A standard remedy is to augment the relative error
criterion and terminate the iteration using a combination of relative and absolute measures of
∇f , i.e., when
(2.12)
∇f (xn )
≤ τr
∇f (x0 )
+ τa .
In (2.12) τa is an absolute error tolerance. Hence, the termination criterion input to many of the
algorithms presented in this book will be in the form of a vector τ = (τr , τa ) of relative and
absolute residuals.
Algorithm 2.3.1. newton(x, f, τ )
1. r0 =
∇f (x)
2. Do while
∇f (x)
> τr r0 + τa
(a) Compute ∇2 f (x)
(b) F actor ∇2 f (x) = LLT
(c) Solve LLT s = −∇f (x)

(d) x = x + s
(e) Compute ∇f (x).
Algorithm newton, as formulated above, is not completely satisfactory. The value of the
objective function f is never used and step 2b will fail if ∇2 f is not positive definite. This failure,
in fact, could serve as a signal that one is too far from a minimizer for Newton’s method to be
directly applicable. However, if we are near enough (see Exercise 2.7.8) to a local minimizer,
as we assume in this chapter, all will be well and we may apply all the results from nonlinear
equations.
2.3.1 Errors in Functions, Gradients, and Hessians

In the presence of errors in functions and gradients, however, the problem of unconstrained
optimization becomes more difficult than that of root finding. We discuss this difference only
briefly here and for the remainder of this chapter assume that gradients are computed exactly, or
at least as accurately as f , say, either analytically or with automatic differentiation [129], [130].
However, we must carefully study the effects of errors in the evaluation of the Hessian just as
we did those of errors in the Jacobian in [154].
A significant difference from the nonlinear equations case arises if only functions are available
and gradients and Hessians must be computed with differences. A simple one-dimensional
analysis will suffice to explain this. Assume that we can only compute f approximately. If we
compute fˆ = f + f rather than f , then a forward difference gradient with difference increment
h
fˆ(x + h) − fˆ(x)
Dh f (x) =
h
√
differs from f by O(h + f /h) and this error is minimized if h = O( f ). In that case the error
√
in the gradient is g = O(h) = O( f ). If a forward difference Hessian is computed using Dh
as an approximation to the gradient, then the error in the Hessian will be
√ 1/4
(2.13) ∆ = O( g ) = O(f )
and the accuracy in ∇2 f will be much less than that of a Jacobian in the nonlinear equations
case.
If f is significantly larger than machine roundoff, (2.13) indicates that using numerical
Hessians based on a second numerical differentiation of the objective function will not be very
accurate. Even in the best possible case, where f is the same size as machine roundoff, finite
difference Hessians will not be very accurate and will be very expensive to compute if the Hessian
is dense. If, as on most computers today, machine roundoff is (roughly) 10−16 , (2.13) indicates
that a forward difference Hessian will be accurate to roughly four decimal digits.
One can obtain better results with centered differences, but at a cost of twice the number of
function evaluations. A centered difference approximation to ∇f is
fˆ(x + h) − fˆ(x − h)
Dh f (x) =
2h
1/3
and the error is O(h2 + f /h), which is minimized if h = O(f ) leading to an error in the
2/3
gradient of g = O(f ). Therefore, a central difference Hessian will have an error of
4/9
∆ = O((g )2/3 ) = O(f ),
which is substantially better. We will find that accurate gradients are much more important than
accurate Hessians and one option is to compute gradients with central differences and Hessians

2/3
with forward differences. If one does that the centered difference gradient error is O(f ) and
therefore the forward difference Hessian error will be
√ 1/3
∆ = O g = O(f ).
More elaborate schemes [22] compute a difference gradient and then reuse the function evalua-
tions in the Hessian computation.
In many optimization problems, however, accurate gradients are available. When that is the
case, numerical differentiation to compute Hessians is, like numerical computation of Jacobians
for nonlinear equations [154], a reasonable idea for many problems and the less expensive
forward differences work well.
Clever implementations of difference computation can exploit sparsity in the Hessian [67],
[59] to evaluate a forward difference approximation with far fewer than N evaluations of ∇f .
In the sparse case it is also possible [22], [23] to reuse the points from a centered difference
approximation to the gradient to create a second-order accurate Hessian.
Unless g (xn ) → 0 as the iteration progresses, one cannot expect convergence. For this
reason estimates like (2.14) are sometimes called local improvement [88] results. Theorem 2.3.4
is a typical example.
Theorem 2.3.4. Let the standard assumptions hold. Then there are K̄ > 0, δ > 0, and
δ1 > 0 such that if xc ∈ B(δ) and
∆(xc )
< δ1 then
x+ = xc − (∇2 f (xc ) + ∆(xc ))−1 (∇f (xc ) + g (xc ))
is defined (i.e., ∇2 f (xc ) + ∆(xc ) is nonsingular) and satisfies
(2.14)
e+
≤ K̄(
ec
2 +
∆(xc )
ec
+
g (xc )
).
Proof. Let δ be small enough so that the conclusions of Lemma 2.3.1 and Theorem 2.3.2
hold. Let
xN 2
+ = xc − (∇ f (xc ))
−1
∇f (xc )
and note that
x+ = xN 2
+ + ((∇ f (xc ))
−1
− (∇2 f (xc ) + ∆(xc ))−1 )∇f (xc ) − (∇2 f (xc ) + ∆(xc ))−1 g (xc ).
Lemma 2.3.1 and Theorem 2.3.2 imply

e+
≤ K
ec
2 + 2
(∇2 f (xc ))−1 − (∇2 f (xc ) + ∆(xc ))−1
∇2 f (x∗ )
ec
(2.15)
+
(∇2 f (xc ) + ∆(xc ))−1
g (xc )
.
If

∆(xc )
≤
(∇2 f (x∗ ))−1
−1 /4,
then Lemma 2.3.1 implies that

∆(xc )
≤
(∇2 f (xc ))−1
−1 /2
and the Banach Lemma [12], [154] states that ∇2 f (xc ) + ∆(xc ) is nonsingular and

(∇2 f (xc ) + ∆(xc ))−1
≤ 2
(∇2 f (xc ))−1
≤ 4
(∇2 f (x∗ ))−1
.
Hence,

(∇2 f (xc ))−1 − (∇2 f (xc ) + ∆(xc ))−1
≤ 8
(∇2 f (x∗ ))−1
2
∆(xc )
.

We use these estimates and (2.15) to obtain
e+
≤ K
ec
2 + 16
(∇2 f (x∗ ))−1
2
∇2 f (x∗ )
∆(xc )
ec
+ 4
(∇2 f (x∗ ))−1
g (xc )
.
Setting
K̄ = K + 16
(∇2 f (x∗ ))−1
2
∇2 f (x∗ )
+ 4
(∇2 f (x∗ ))−1
completes the proof.

As is the case with equations, (2.14) implies that one cannot hope to find a minimizer with
more accuracy that one can evaluate ∇f . In most cases the iteration will stagnate once
e
is
(roughly) the same size as g . The speed of convergence will be governed by the accuracy in the
Hessian.
The result for the chord method illustrates this latter point. In the chord method we form
and compute the Cholesky factorization of ∇2 f (x0 ) and use that factorization to compute all
subsequent Newton steps. Hence,
x+ = xc − (∇2 f (x0 ))−1 ∇f (xc )
and
(2.16)
∆(xc )
≤ γ
x0 − xc
≤ γ(
e0
+
ec
).
Algorithmically the chord iteration differs from the Newton iteration only in that the computation
and factorization of the Hessian is moved outside of the main loop.
Algorithm 2.3.2. chord(x, f, τ )
1. r0 =
∇f (x)
2. Compute ∇2 f (x)
3. F actor ∇2 f (x) = LLT
4. Do while
∇f (x)
> τr r0 + τa
(a) Solve LLT s = −∇f (x)
(b) x = x + s
(c) Compute ∇f (x).
The convergence theory follows from Theorem 2.3.4 with g = 0 and ∆ = O(

e0
).
Theorem 2.3.5. 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
(2.17)
en+1
≤ KC
e0
en
.
Proof. Let δ be small enough so that the conclusions of Theorem 2.3.4 hold. Assume that
en , e0 ∈ B(δ). Combining (2.16) and (2.14) implies

en+1
≤ K̄(
en
(1 + γ) + γ
e0
)
en
≤ K̄(1 + 2γ)δ
en
.
Hence, if δ is small enough so that
K̄(1 + 2γ)δ = η < 1,
then the chord iterates converge q-linearly to x∗ . Q-linear convergence implies that
en
<
e0
and hence (2.17) holds with KC = K̄(1 + 2γ).

The Shamanskii method [233], [154], [211] is a generalization of the chord method that
updates Hessians after every m + 1 nonlinear iterations. Newton’s method corresponds to
m = 1 and the chord method to m = ∞. The convergence result is a direct consequence of
Theorems 2.3.3 and 2.3.5.
Theorem 2.3.6. Let the standard assumptions hold and let m ≥ 1 be given. 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 and
(2.18)
en+1
≤ KS
en
m+1 .
As one more application of Theorem 2.3.4, we analyze the effects of a difference approxima-
tion of the Hessian. We follow the notation of [154] where possible. For example, to construct
a Hessian matrix, whose columns are ∇2 f (x)ej , where ej is the unit vector with jth compo-
nent 1 and other components 0, we could approximate the matrix–vector products ∇2 f (x)ej by
forward differences and then symmetrize the resulting matrix. We define
(2.19) ∇2h f (x) = (Ah + ATh )/2,
where Ah is the matrix whose jth column is Dh2 f (x : ej ). Dh2 f (x : w), the difference approxi-
mation of the action of the Hessian ∇2 f (x) on a vector w, is defined to be the quotient


 0, w = 0,

2
(2.20) Dh f (x : w) =

 ∇f (x + hw/
w
) − ∇f (x)
 , w = 0.
h/
w
Note that we may also write

Dh2 f (x : w) = Dh (∇f )(x : w),
where the notation Dh , taken from [154], denotes numerical directional derivative. If
x
is very
large, then the error in computing the sum x + hw/
w
will have to be taken into consideration
in the choice of h.
We warn the reader, as we did in [154], that D2 f (x : w) is not a linear map and that
D f (x : w) is not, in general, the same as ∇2h f (x)w.
2
If we compute ∇f (x) + g (x) and the gradient errors satisfy an estimate of the form

g (x)
≤ ¯
for all x, then the computed difference Hessian is ∇h (∇f + g ) and satisfies
(2.21)
∇2 f (x) − ∇h (∇f + g )(x)
= O(h + ¯/h).
√
So, as in [154], the choice h ≈ ¯ is optimal in the sense that it minimizes the quantity in the
O-term in (2.21).
The local convergence theorem in this case is [88], [154], [278], as follows.
Theorem 2.3.7. Let the standard assumptions hold. Then there are δ, ¯, and KD > 0 such
that if xc ∈ B(δ),
g (x)
≤ ¯c for all x ∈ B(δ), and

h ≥ M
g (xc )
then
x+ = xc − (∇hc (∇f (xc ) + g (xc )))−1 (∇f (xc ) + g (xc ))
satisfies

e+
≤ KD (¯
c + (¯
c + h)
ec
).

2.3.2 Termination of the Iteration

It is not safe to terminate the iteration when f (xc ) − f (x+ ) is small, and no conclusions can
safely be drawn by examination of the differences of the objective function values at successive
iterations. While some of the algorithms for difficult problems in Part II of this book do indeed
terminate when successive function values are close, this is an act of desperation. For example,
if
n

f (xn ) = − j −1 ,
j=1
then f (xn ) → −∞ but f (xn+1 ) − f (xn ) = −1/(n + 1) → 0. The reader has been warned.
If the standard assumptions hold, then one may terminate the iteration when the norm of ∇f
is sufficiently small relative to ∇f (x0 ) (see [154]). We will summarize the key points here and
refer the reader to [154] for the details. The idea is that if ∇2 f (x∗ ) is well conditioned, then a
small gradient norm implies a small error norm. Hence, for any gradient-based iterative method,
termination on small gradients is reasonable. In the special case of Newton’s method, the norm
of the step is a very good indicator of the error and if one is willing to incur the added cost of an
extra iteration, a very sharp bound on the error can be obtained, as we will see below.
Lemma 2.3.8. Assume that the standard assumptions hold. Let δ > 0 be small enough so
that the conclusions of Lemma 2.3.1 hold for x ∈ B(δ). Then for all x ∈ B(δ)

e

∇f (x)
4κ(∇2 f (x∗ ))
e
(2.22) ≤ ≤ .
4
e0
κ(∇2 f (x∗ ))
∇f (x0 )

e0
The meaning of (2.22) is that, up to a constant multiplier, the norm of the relative gradient
is the same as the norm of the relative error. This partially motivates the termination condition
(2.12).
In the special case of Newton’s method, one can use the steplength as an accurate estimate
of the error because Theorem 2.3.2 implies that
(2.23)
ec
=
s
+ O(
ec
2 ).
Hence, near the solution s and ec are essentially the same size. The cost of using (2.23) is that
all the information needed to compute x+ = xc + s has been computed. If we terminate the
iteration when
s
is smaller than our desired tolerance and then take x+ as the final result, we
have attained more accuracy than we asked for. One possibility is to terminate the iteration when
√ √
s
= O( τs ) for some τs > 0. This, together with (2.23), will imply that
ec
= O( τs )
and hence, using Theorem 2.3.2, that
(2.24)
e+
= O(
ec
2 ) = O(τs ).
For a superlinearly convergent method, termination on small steps is equally valid but one
cannot use (2.24). For a superlinearly convergent method we have
(2.25)
ec
=
s
+ o(
ec
) and
e+
= o(
ec
).
Hence, we can conclude that

e+
< τs if
s
< τs . This is a weaker, but still very useful,
result.
For a q-linearly convergent method, such as the chord method, making termination decisions
based on the norms of the steps is much riskier. The relative error in estimating
ec
by
s
is
|
ec
−
s
|
ec + s

e+
≤ = .

ec

ec

ec

Hence, estimation of errors by steps is worthwhile only if convergence is fast. One can go further
[156] if one has an estimate ρ of the q-factor that satisfies

e+
≤ ρ
ec
.
In that case,
(1 − ρ)
ec
≤
ec
−
e+
≤
ec − e+
=
s
.
Hence
ρ
(2.26)
e+
≤ ρ
ec
≤
s
.
1−ρ
So, if the q-factor can be estimated from above by ρ and

s
< (1 − ρ)τs /ρ,
then
e+
< τs . This approach is used in ODE and DAE codes [32], [234], [228], [213],
but requires good estimates of the q-factor and we do not advocate it for q-linearly convergent
methods for optimization. The danger is that if the convergence is slow, the approximate q-factor
can be a gross underestimate and cause premature termination of the iteration.
It is not uncommon for evaluations of f and ∇f to be very expensive and optimizations are,
therefore, usually allocated a fixed maximum number of iterations. Some algorithms, such as
the DIRECT, [150], algorithm we discuss in §8.4.2, assign a limit to the number of function
evaluations and terminate the iteration in only this way.
2.4 Nonlinear Least Squares

Nonlinear least squares problems have objective functions of the form
M
1 1
(2.27) f (x) =
ri (x)
22 = R(x)T R(x).
2 i=1 2
The vector R = (r1 , . . . , rM ) is called the residual. These problems arise in data fitting, for
example. In that case M is the number of observations and N is the number of parameters;
for these problems M > N and we say the problem is overdetermined. If M = N we have a
nonlinear equation and the theory and methods from [154] are applicable. If M < N the problem
is underdetermined. Overdetermined least squares problems arise most often in data fitting
applications like the parameter identification example in §1.6.2. Underdetermined problems are
less common, but are, for example, important in the solution of high-index differential algebraic
equations [48], [50].
The local convergence theory for underdetermined problems has the additional complexity
that the limit of the Gauss–Newton iteration is not uniquely determined by the distance of the
initial iterate to the set of points where R(x∗ ) = 0. In §2.4.3 we describe the difficulties and
state a simple convergence result.
If x∗ is a local minimizer of f and f (x∗ ) = 0, the problem min f is called a zero residual
problem (a remarkable and suspicious event in the data fitting scenario). If f (x∗ ) is small, the
expected result in data fitting if the model (i.e., R) is good, the problem is called a small residual
problem. Otherwise one has a large residual problem.
Nonlinear least squares problems are an intermediate stage between nonlinear equations and
optimization problems and the methods for their solution reflect this. We define the M × N
Jacobian R of R by
(2.28) (R (x))ij = ∂ri /∂xj , 1 ≤ i ≤ M, 1 ≤ j ≤ N.

With this notation it is easy to show that
(2.29) ∇f (x) = R (x)T R(x) ∈ RN .
The necessary conditions for optimality imply that at a minimizer x∗
(2.30) R (x∗ )T R(x∗ ) = 0.
In the underdetermined case, if R (x∗ ) has full row rank, (2.30) implies that R(x∗ ) = 0; this is
not the case for overdetermined problems.
The cost of a gradient is roughly that of a Jacobian evaluation, which, as is the case with
nonlinear equations, is the most one is willing to accept. Computation of the Hessian (an N × N
matrix)
M
(2.31) ∇2 f (x) = R (x)T R (x) + ri (x)T ∇2 ri (x)
j=1
2
requires computation of the M Hessians {∇ ri } for the second-order term
M
ri (x)T ∇2 ri (x)
j=1
and is too costly to be practical.

We will also express the second-order term as
M
ri (x)T ∇2 ri (x) = R (x)T R(x),
j=1
where the second derivative R is a tensor. The notation is to be interpreted in the following
way. For v ∈ RM , R (x)T v is the N × N matrix
M
R (x)T v = ∇2 (R(x)T v) = (v)i ∇2 ri (x).
i=1
We will use the tensor notation when expanding R about x∗ in some of the analysis to follow.
2.4.1 Gauss–Newton Iteration

The Gauss–Newton algorithm simply discards the second-order term in ∇2 f and computes a
step
s = −(R (xc )T R (xc ))−1 ∇f (xc )
(2.32)
= −(R (xc )T R (xc ))−1 R (xc )T R(xc ).
The Gauss–Newton iterate is x+ = xc +s. One motivation for this approach is that R (x)T R(x)
vanishes for zero residual problems and therefore might be negligible for small residual problems.
Implicit in (2.32) is the assumption that R (xc )T R (xc ) is nonsingular, which implies that
M ≥ N . Another interpretation, which also covers underdetermined problems, is to say that the
Gauss–Newton iterate is the minimum norm solution of the local linear model of our nonlinear
least squares problem
1
(2.33) min
R(xc ) + R (xc )(x − xc )
2 .
2
Using (2.33) and linear least squares methods is a more accurate way to compute the step than
using (2.32), [115], [116], [127]. In the underdetermined case, the Gauss–Newton step can
be computed with the singular value decomposition [49], [127], [249]. (2.33) is an overde-
termined, square, or underdetermined linear least squares problem if the nonlinear problem is
overdetermined, square, or underdetermined.

The standard assumptions for nonlinear least squares problems follow in Assumption 2.4.1.
Assumption 2.4.1. x∗ is a minimizer of
R
22 , R is Lipschitz continuously differentiable
near x∗ , and R (x∗ )T R (x∗ ) has maximal rank. The rank assumption may also be stated as
• R (x∗ ) is nonsingular (M = N ),
• R (x∗ ) has f ull column rank (M > N ),
• R (x∗ ) has f ull row rank (M < N ).
2.4.2 Overdetermined Problems
Theorem 2.4.1. Let M > N . Let Assumption 2.4.1 hold. Then there are K > 0 and δ > 0
such that if xc ∈ B(δ) then the error in the Gauss–Newton iteration satisfies
(2.34)
e+
≤ K(
ec
2 +
R(x∗ )
ec
).
Proof. Let δ be small enough so that

x − x∗
< δ implies that R (x)T R (x) is nonsingular.
Let γ be the Lipschitz constant for R .
By (2.32)
e+ = ec − (R (xc )T R (xc ))−1 R (xc )T R(xc )

(2.35)
= (R (xc )T R (xc ))−1 R (xc )T (R (xc )ec − R(xc )).
Note that
R (xc )ec − R(xc ) = R (xc )ec − R(x∗ ) + R(x∗ ) − R(xc )
= −R(x∗ ) + (R (xc )ec + R(x∗ ) − R(xc )).
Now,

R (xc )ec + R(x∗ ) − R(xc )
≤ γ
ec
2 /2
and, since R (x∗ )T R(x∗ ) = 0,
−R (xc )T R(x∗ ) = (R (x∗ ) − R (xc ))T R(x∗ ).
Hence,

e+
≤
(R (xc )T R (xc ))−1
(R (x∗ ) − R (xc ))T R(x∗ )

(R (xc )T R (xc ))−1
R (xc )T
γ
ec
2
+
(2.36) 2

T −1
R(x∗ )
+
R (xc )T
ec
≤
(R (xc ) R (xc ))
γ
ec
.
2
Setting
1 +
R (x)T
K = γ max (
(R (x)T R (x))−1
x∈B(δ) 2

completes the proof.

There are several important consequences of Theorem 2.4.1. The first is that for zero residual
problems, the local convergence rate is q-quadratic because the
R(x∗ )
ec
term on the right
side of (2.34) vanishes. For a problem other than a zero residual one, even q-linear convergence
is not guaranteed. In fact, if xc ∈ B(δ) then (2.35) will imply that
e+
≤ r
ec
for some
0 < r < 1 if
(2.37) K(δ +
R (x∗ )
) ≤ r
and therefore the q-factor will be K
R (x∗ )
. Hence, for small residual problems and accurate
initial data the convergence of Gauss–Newton will be fast. Gauss–Newton may not converge at
all for large residual problems.
Equation (2.36) exposes a more subtle issue when the term
(R (x∗ ) − R (xc ))T R(x∗ )
is considered as a whole, rather than estimated by
γ
ec
R(x∗ )
.
Using Taylor’s theorem and the necessary conditions (R (x∗ )T R(x∗ ) = 0) we have
R (xc )T R(x∗ ) = [R (x∗ ) + R (x∗ )ec + O(

ec
2 )]T R(x∗ )
= eTc R (x∗ )T R(x∗ ) + O(

ec
2 ).
Recall that
R (x∗ )T R(x∗ ) = ∇2 f (x∗ ) − R (x∗ )T R (x∗ )
and hence

(R (x∗ )−R (xc ))T R(x∗ )
(2.38)
≤
∇2 f (x∗ ) − R (x∗ )T R (x∗ )
R(x∗ )
+ O(
ec
2 ).
In a sense (2.38) says that even for a large residual problem, convergence can be fast if the problem
is not very nonlinear (small R ). In the special case of a linear least squares problem (where
R = 0) Gauss–Newton becomes simply the solution of the normal equations and converges in
one iteration.
So, Gauss–Newton can be expected to work well for overdetermined small residual problems
and good initial iterates. For large residual problems and/or initial data far from the solution,
there is no reason to expect Gauss–Newton to give good results. We address these issues in
§3.2.3.
2.4.3 Underdetermined Problems

We begin with the linear underdetermined least squares problem
(2.39) min
Ax − b
2 .
If A is M × N with M < N there will not be a unique minimizer but there will be a unique
minimizer with minimum norm. The minimum norm solution can be expressed in terms of the
singular value decomposition [127], [249] of A,
(2.40) A = U ΣV T .

In (2.40), Σ is an N × N diagonal matrix. The diagonal entries of Σ, {σi } are called the singular
values. σi ≥ 0 and σi = 0 if i > M . The columns of the M × N matrix U and the N × N
matrix V are called the left and right singular vectors. U and V have orthonormal columns and
hence the minimum norm solution of (2.39) is
x = A† b,
where A† = V Σ† U T ,
Σ† = diag(σ1† , . . . , σN
†
),
and  −1
 σi , σi = 0,
σi† =

0, σi = 0.
A† is called the Moore–Penrose inverse [49], [189], [212]. If A is a square nonsingular matrix,
then A† = A−1 ; if M > N then the definition of A† using the singular value decomposition is
still valid; and, if A has full column rank, A† = (AT A)−1 AT .
Two simple properties of the Moore–Penrose inverse are that A† A is a projection onto the
range of A† and AA† is a projection onto the range of A. This means that
(2.41) A† AA† = A† , (A† A)T = A† A, AA† A = A, and (AA† )T = AA† .
So the minimum norm solution of the local linear model (2.33) of an underdetermined
nonlinear least squares problem can be written as [17], [102]
(2.42) s = −R (xc )† R(xc )
and the Gauss–Newton iteration [17] is
(2.43) x+ = xc − R (xc )† R(xc ).
The challenge in formulating a local convergence result is that there is not a unique optimal point
that attracts the iterates.
In the linear case, where R(x) = Ax − b, one gets
x+ = xc − A† (Axc − b) = (I − A† A)xc − A† b.
Set e = x − A† b and note that

A† AA† b = A† b
by (2.41). Hence
e+ = (I − A† A)ec .
This does not imply that x+ = A† b, the minimum norm solution, only that x+ is a solution of
the problem and the iteration converges in one step. The Gauss–Newton iteration cannot correct
for errors that are not in the range of A† .
Let
Z = {x | R(x) = 0}.
We show in Theorem 2.4.2, a special case of the result in [92], that if the standard assumptions
hold at a point x∗ ∈ Z, then the iteration will converge q-quadratically to a point z ∗ ∈ Z.
However, there is no reason to expect that z ∗ = x∗ . In general z ∗ will depend on x0 , a very
different situation from the overdetermined case. The hypotheses of Theorem 2.4.2, especially
that of full column rank in R (x), are less general than those in [24], [17], [25], [92], and [90].

Theorem 2.4.2. Let M ≤ N and let Assumption 2.4.1 hold for some x∗ ∈ Z. Then there is
δ > 0 such that if

x0 − x∗
≤ δ,
then the Gauss–Newton iterates
xn+1 = xn − R (xn )† R(xn )
exist and converge r-quadratically to a point z ∗ ∈ Z.

Proof. Assumption 2.4.1 and results in [49], [126] imply that if δ is sufficiently small then
there is ρ1 such that R (x)† is Lipschitz continuous in the set
B1 = {x |
x − x∗
≤ ρ1 }
and the singular values of R (x) are bounded away from zero in B1 . We may, reducing ρ1 if
necessary, apply the Kantorovich theorem [154], [151], [211] to show that if x ∈ B1 and w ∈ Z
is such that

x − w
= min
x − z
,
z∈Z
then there is ξ = ξ(x) ∈ Z such that

w − ξ(x)
= O(
x − w
2 ) ≤
x − w
/2
and ξ is in the range of R (w)† , i.e.,
R (w)† R (w)(x − ξ(x)) = x − ξ(x).
The method of the proof is to adjust δ so that the Gauss–Newton iterates remain in B1 and
R(xn ) → 0 sufficiently rapidly. We begin by requiring that δ < ρ1 /2.
Let xc ∈ B1 and let e = x − ξ(xc ). Taylor’s theorem, the fundamental theorem of calculus,
and (2.41) imply that
e+ = ec − R (xc )† R(xc )
= ec − (R (xc )† − R (ξ(x))† )R(x) − R (x∗)† R(x)
= e0 − R (x∗ )† R(x) + O(
ec
2 )
= (I − R (x∗ )† R (x∗ ))ec + O(

ec
2 ) = O(
ec
2 ).
If we define d(x) = minz∈Z

x − z
then there is K1 such that
(2.44) d(x+ ) ≤
x+ − ξ(xc )
≤ K1
xc − ξ(xc )
2 ≤ K1 d(xc )2 .
Now let
ρ2 = min(ρ1 , (2K1 )−1 ).
So if
xc ∈ B2 = {x |
x − x∗
≤ ρ2 }
then
(2.45) d(x+ ) ≤ d(xc )/2.
Finally, there is K2 such that

x+ − x∗
≤
xc − x∗
+
x+ − xc
=
xc − x∗
+
R (xc )† R(xc )
≤
xc − x∗
+ K2
xc − ξ(xc )
.

We now require that

ρ2
(2.46) δ≤ .
2(1 + K2 )
We complete the proof by induction. If
x0 − x∗
≤ δ and the Gauss–Newton iterates {xk }nk=0
are in B2 , then xn+1 is be defined and, using (2.46) and (2.44),
n+1

xn+1 − x∗
≤
x0 − x∗
+ K3 d(xk ) ≤ δ + 2K3 d(x0 ) ≤ ρ1 .
k=0
Hence, the Gauss–Newton iterates exist, remain in B0 , and dn → 0.

To show that the sequence of Gauss–Newton iterates does in fact converge, we observe that
there is K3 such that

x+ − xc
=
R (xc )† R(xc )
≤ K3
xc − ξ(xc )
≤ K3 d(xc ).
Therefore (2.45) implies that for any m, n ≥ 0,

n+m−1

xn+m − xn
≤ l=n
xl+1 − xl
n+m 1 − 2−m
= l=n d(xl ) = d(xn )
2
≤ 2d(xn ) ≤ 2−n+1 d(x0 ).
Hence, {xk } is a Cauchy sequence and therefore converges to a point z ∗ ∈ Z. Since

xn − z ∗
≤ 2d(xn ),
(2.44) implies that the convergence is r-quadratic.
2.5 Inexact Newton Methods

An inexact Newton method [74] uses an approximate Newton step s = x+ − xc , requiring only
that
(2.47)
∇2 f (xc )s + ∇f (xc )
≤ ηc
∇f (xc )
,
i.e., that the linear residual be small. We will refer to any vector s that satisfies (2.47) with
ηc < 1 as an inexact Newton step. We will refer to the parameter ηc on the right-hand side of
(2.47) as the forcing term [99] .
Inexact Newton methods are also called truncated Newton methods [75], [198], [199] in the
context of optimization. In this book, we consider Newton–iterative methods. This is the class of
inexact Newton methods in which the linear equation (2.4) for the Newton step is also solved by
an iterative method and (2.47) is the termination criterion for that linear iteration. It is standard
to refer to the sequence of Newton steps {xn } as the outer iteration and the sequence of iterates
for the linear equation as the inner iteration. The naming convention (see [33], [154], [211])
is that Newton–CG, for example, refers to the Newton–iterative method in which the conjugate
gradient [141] algorithm is used to perform the inner iteration.
Newton–CG is particularly appropriate for optimization, as we expect positive definite Hes-
sians near a local minimizer. The results for inexact Newton methods from [74] and [154]
are sufficient to describe the local convergence behavior of Newton–CG, and we summarize
the relevant results from nonlinear equations in §2.5.1. We will discuss the implementation of
Newton–CG in §2.5.2.

2.5.1 Convergence Rates

We will prove the simplest of the convergence results for Newton–CG, Theorem 2.5.1, from
which Theorem 2.5.2 follows directly. We refer the reader to [74] and [154] for the proof of
Theorem 2.5.3.
Theorem 2.5.1. Let the standard assumptions hold. Then there are δ and KI such that if
xc ∈ B(δ), s satisfies (2.47), and x+ = xc + s, then
(2.48)
e+
≤ KI (
ec
+ ηc )
ec
.
Proof. Let δ be small enough so that the conclusions of Lemma 2.3.1 and Theorem 2.3.2
hold. To prove the first assertion (2.48) note that if
r = −∇2 f (xc )s − ∇f (xc )
is the linear residual, then
s + (∇2 f (xc ))−1 ∇f (xc ) = −(∇2 f (xc ))−1 r
and
(2.49) e+ = ec + s = ec − (∇2 f (xc ))−1 ∇f (xc ) − (∇2 f (xc ))−1 r.
Now, (2.47), (2.7), and (2.6) imply that

s + (∇2 f (xc ))−1 ∇f (xc )
≤
(∇2 f (xc ))−1
ηc
∇f (xc )
≤ 4κ(∇2 f (x∗ ))ηc

ec
.
Hence, using (2.49) and Theorem 2.3.2, we have that

e+
≤
ec − ∇2 f (xc )−1 ∇f (xc )
+ 4κ(F (x∗ ))ηc
ec
≤ K
ec
2 + 4κ(∇2 f (x∗ ))ηc
ec
,
where K is the constant from (2.8). If we set
KI = K + 4κ(∇2 f (x∗ )),
the proof is complete.

Theorem 2.5.2. Let the standard assumptions hold. Then there are δ and η̄ such that if
x0 ∈ B(δ), {ηn } ⊂ [0, η̄], then the inexact Newton iteration
xn+1 = xn + sn ,
where

∇2 f (xn )sn + ∇f (xn )
≤ ηn
∇f (xn )
,
converges q-linearly to x∗ . Moreover
• if ηn → 0 the convergence is q-superlinear, and
• if ηn ≤ Kη
∇f (xn )
p for some Kη > 0 the convergence is q-superlinear with q-order
1 + p.

The similarity between Theorem 2.5.2 and Theorem 2.3.5, the convergence result for the
chord method, should be clear. Rather than require that the approximate Hessian be accurate,
we demand that the linear iteration produce a sufficiently small relative residual. Theorem 2.5.3
is the remarkable statement that any reduction in the relative linear residual will suffice for linear
convergence in a certain norm. This statement implies [154] that
∇f (xn )
will converge to
zero q-linearly, or, equivalently, that xn → x∗ q-linearly with respect to
·
∗ , which is defined
by

x
∗ =
∇2 f (x∗ )x
.
Theorem 2.5.3. Let the standard assumptions hold. Then there is δ such that if xc ∈ B(δ),
s satisfies (2.47), x+ = xc + s, and ηc ≤ η < η̄ < 1, then
(2.50)
e+
∗ ≤ η̄
ec
∗ .
Theorem 2.5.4. Let the standard assumptions hold. Then there is δ such that if x0 ∈ B(δ),
{ηn } ⊂ [0, η] with η < η̄ < 1, then the inexact Newton iteration
xn+1 = xn + sn ,
where

∇2 f (xn )sn + ∇f (xn )
≤ ηn
∇f (xn )
converges q-linearly with respect to

·
∗ to x∗ . Moreover
• if ηn ≤ Kη
∇f (xn )
1 + p.
Q-linear convergence of {xn } to a local minimizer with respect to

·
∗ is equivalent to
q-linear convergence of {∇f (xn )} to zero. We will use the rate of convergence of {∇f (xn )}
in our computational examples to compare various methods.
2.5.2 Implementation of Newton–CG

Our implementation of Newton–CG approximately solves the equation for the Newton step with
CG. We make the implicit assumption that ∇f has been computed sufficiently accurately for
Dh2 f (x : w) to be a useful approximate Hessian of the Hessian–vector product ∇2 f (x)w.
Forward Difference CG
Algorithm fdcg is an implementation of the solution by CG of the equation for the Newton step
(2.4). In this algorithm we take care to identify failure in CG (i.e., detection of a vector p for
which pT Hp ≤ 0). This failure either means that H is singular (pT Hp = 0; see exercise 2.7.3)
or that pT Hp < 0, i.e., p is a direction of negative curvature. The algorithms we will discuss
in §3.3.7 make good use of directions of negative curvature. The initial iterate for forward
difference CG iteration should be the zero vector. In this way the first iterate will give a steepest
descent step, a fact that is very useful. The inputs to Algorithm fdcg are the current point x,
the objective f , the forcing term η, and a limit on the number of iterations kmax. The output is
the inexact Newton direction d. Note that in step 2b Dh2 f (x : p) is used as an approximation to
∇2 f (x)p.
Algorithm 2.5.1. fdcg(x, f, η, kmax, d)

1. r = −∇f (x), ρ0 =
r
22 , k = 1, d = 0.
√
2. Do While ρk−1 > η
∇f (x)
and k < kmax
(a) if k = 1 then p = r
else
β = ρk−1 /ρk−2 and p = r + βp
(b) w = Dh2 f (x : p)
If pT w = 0 signal indefiniteness; stop.
If pT w < 0 signal negative curvature; stop.
(c) α = ρk−1 /pT w
(d) d = d + αp
(e) r = r − αw
(f) ρk =
r
2
(g) k = k + 1
Preconditioning can be incorporated into a Newton–CG algorithm by using a forward dif-

ference formulation, too. Here, as in [154], we denote the preconditioner by M . Aside from M ,
the inputs and output of Algorithm fdpcg are the same as that for Algorithm fdcg.
Algorithm 2.5.2. fdpcg(x, f, M, η, kmax, d)
1. r = −∇f (x), ρ0 =
r
22 , k = 1, d = 0.
√
∇f (x)
and k < kmax
(a) z = M r
(b) τk−1 = z T r
else
β = τk−1 /τk−2 , p = z + βp
(d) w = Dh2 f (x : p)
If pT w = 0 signal indefiniteness; stop.
If pT w < 0 signal negative curvature; stop.
(e) α = τk−1 /pT w
(f) d = d + αp
(g) r = r − αw
(h) ρk = rT r
(i) k = k + 1
In our formulation of Algorithms fdcg and fdpcg, indefiniteness is a signal that we are
not sufficiently near a minimum for the theory in this section to hold. In §3.3.7 we show how
negative curvature can be exploited when far from the solution.
One view of preconditioning is that it is no more than a rescaling of the independent variables.
Suppose, rather than (1.2), we seek to solve
(2.51) min fˆ(y),

y

where fˆ(y) = f (M 1/2 y) and M is spd. If y ∗ is a local minimizer of fˆ, then x∗ = M 1/2 y ∗ is
a local minimizer of f and the two problems are equivalent. Moreover, if x = M 1/2 y and ∇x
and ∇y denote gradients in the x and y coordinates, then
∇y fˆ(y) = M 1/2 ∇x f (x)
and
∇2y fˆ(y) = M 1/2 (∇2x f (x))M 1/2 .
Hence, the scaling matrix plays the role of the square root of the preconditioner for the precon-
ditioned conjugate gradient algorithm.
Newton–CG
The theory guarantees that if x0 is near enough to a local minimizer then ∇2 f (xn ) will be spd
for the entire iteration and xn will converge rapidly to x∗ . Hence, Algorithm newtcg will not
terminate with failure because of an increase in f or an indefinite Hessian. Note that both the
forcing term η and the preconditioner M can change as the iteration progresses.
Algorithm 2.5.3. newtcg(x, f, τ, η)
1. rc = r0 =
∇f (x)
2. Do while
∇f (x)
> τr r0 + τa
(a) Select η and a preconditioner M .

(b) fdpcg(x, f, M, η, kmax, d)
If indefiniteness has been detected, terminate with failure.
(c) x = x + d.
(d) Evaluate f and ∇f (x).
If f has not decreased, terminate with failure.
(e) r+ =
∇f (x)
, σ = r+ /rc , rc = r+ .
The implementation of Newton–CG is simple, but, as presented in Algorithm newtcg,

incomplete. The algorithm requires substantial modification to be able to generate the good
initial data that the local theory requires. We return to this issue in §3.3.7.
There is a subtle problem with Algorithm fdpcg in that the algorithm is equivalent to the
application of the preconditioned conjugate gradient algorithm to the matrix B that is determined
by
Bpi = wi = Dh2 f (x : pi ), 1 ≤ i ≤ N.
However, since the map p → Dh2 f (x : p) is not linear in p, the quality of B as an approximation
to ∇2 f (x) may degrade as the linear iteration progresses. Usually this will not cause problems
unless many iterations are needed to satisfy the inexact Newton condition. However, if one does
not see the expected rate of convergence in a Newton–CG iteration, this could be a factor [128].
One partial remedy is to use a centered-difference Hessian–vector product [162], which reduces
the error in B. In exercise 2.7.15 we discuss a more complex and imaginative way to compute
accurate Hessians.

2.6 Examples
In this section we used the collection of MATLAB codes but disabled the features (see Chapter 3)
that assist in convergence when far from a minimizer. We took care to make certain that the
initial iterates were near enough to the minimizer so that the observations of local convergence
corresponded to the theory. In practical optimization problems, good initial data is usually not
available and the globally convergent methods discussed in Chapter 3 must be used to start the
iteration.
The plots in this section have the characteristics of local convergence in that both the gradient
norms and function values are decreasing. The reader should contrast this with the examples in
Chapter 3.

For this example, M = 100, and the observations are those of the exact solution with c = k = 1,
which we computed analytically. We used T = 10 and u0 = 10. We computed the displacement
and solved the sensitivity equations with the stiff solver ode15s. These results could be
obtained as well in a FORTRAN environment using, for example, the LSODE code [228]. The
relative and absolute error tolerances for the integrator were both set to 10−8 . In view of the
expected accuracy of the gradient, we set the forward difference increment for the approximate
Hessian to h = 10−4 . We terminated the iterations when
∇f
< 10−4 . Our reasons for this
are that, for the zero residual problem considered here, the standard assumptions imply that
f (x) = O(
∇f (x)
) for x near the solution. Hence, since we can only resolve f to an accuracy
of 10−8 , iteration beyond the point where
∇f
< 10−4 cannot be expected to lead to a further
decrease in f . In fact we observed this in our computations.
The iterations are very sensitive to the initial iterate. We used x0 = (1.1, 1.05)T ; initial
iterates much worse than that caused Newton’s method to fail. The more robust methods from
Chapter 3 should be viewed as essential components of even a simple optimization code.
In Table 2.1 we tabulate the history of the iteration for both the Newton and Gauss–Newton
methods. As expected for a small residual problem, Gauss–Newton performs well and, for this
example, even converges in fewer iterations. The real benefit of Gauss–Newton is that com-
putation of the Hessian can be avoided, saving considerable computational work by exploiting
the structure of the problem. In the computation reported here, the MATLAB flops com-
mand indicates that the Newton iteration took roughly 1.9 million floating point operations and
Gauss–Newton roughly 650 thousand. This difference would be much more dramatic if there
were more than two parameters or the cost of an evaluation of f depended on N in a significant
way (which it does not in this example).
Table 2.1: Parameter identification problem, locally convergent iterations.
Newton Gauss–Newton
n
∇f (xn )
f (xn )
∇f (xn )
f (xn )
0 2.33e+01 7.88e-01 2.33e+01 7.88e-01
1 6.87e+00 9.90e-02 1.77e+00 6.76e-03
2 4.59e-01 6.58e-04 1.01e-02 4.57e-07
3 2.96e-03 3.06e-08 9.84e-07 2.28e-14
4 2.16e-06 4.15e-14
Figure 2.1 is a graphical representation of the convergence history from Table 2.1. We think
that the plots are a more effective way to understand iteration statistics and will present mostly

graphs for the remainder of the book. The concavity of the plots of the gradient norms is the
signature of superlinear convergence.
Newton’s Method Newton’s Method

2 0
10 10
0
10
Function Value
Gradient Norm
−5
10
−2
10
−10
−4
10
10
−6 −15
10 10
0 2 4 0 2 4
Iterations Iterations
Gauss−Newton Method Gauss−Newton Method
2 0
10 10
0
10
Function Value
Gradient Norm
−5
−2 10
10
−4
10 −10
10
−6
10
−8 −15
10 10
0 1 2 3 0 1 2 3
Figure 2.1: Local Optimization for the Parameter ID Problem
We next illustrate the difference between Gauss–Newton and Newton on a nonzero residual
problem. We use the same example as before with the observations randomly perturbed. We
used the MATLAB rand function for this, perturbing the samples of the analytic solution by
.5 × rand(M, 1). The least squares residual is about 3.6 and the plots in Figure 2.2 indicate
that Newton’s method is still converging quadratically, but the rate of Gauss–Newton is linear.
The linear convergence of Gauss–Newton can be seen clearly from the linear semilog plot of the
gradient norms. Even so, the Gauss–Newton iteration was more efficient, in terms of floating
point operation, than Newton’s method. The Gauss–Newton iteration took roughly 1 million
floating point operations while the Newton iteration took 1.4 million.
2.6.2 Discrete Control Problem

We solve the discrete control problem from §1.6.1 with N = 400, T = 1, y0 = 0,
L(y, u, t) = (y − 3)2 + .5u2 , and φ(y, u, t) = uy + t2
with Newton–CG and two different choices, η = .1, .0001, of the forcing term. The initial
iterate was u0 = (10, 10, . . . , 10)T and the iteration was terminated when
∇f
< 10−8 . In

Newton’s Method Newton’s Method

2
10 4
0
10 3.9
Function Value
Gradient Norm
−2
10 3.8
−4
10 3.7
−6
10 3.6
0 1 2 3 0 1 2 3
Gauss−Newton Method Gauss−Newton Method
2
10 4
0
10 3.9
Function Value
Gradient Norm
−2
10 3.8
−4
10 3.7
−6
10 3.6
0 2 4 6 0 2 4 6
Figure 2.2: Local Optimization for the Parameter ID Problem, Nonzero Residual
Figure 2.3 one can see that the small forcing term produces an iteration history with the concavity
of superlinear convergence. The limiting q-linear behavior of an iteration with constant η is not
yet visible. The iteration with the larger value of η is in the q-linearly convergent stage, as the
linear plot of ∇f against the iteration counter shows.
The cost of the computation is not reflected by the number of nonlinear iterations. When
η = .0001, the high accuracy of the linear solve is not rewarded. The computation with η = .0001
required 8 nonlinear iterations, a total of 32 CG iterations, roughly 1.25 million floating point
operations, and 41 gradient evaluations. The optimization with η = .1 needed 10 nonlinear
iterations, a total of 13 CG iterations, roughly 820 thousand floating point operations, and 24
gradient evaluations.
2.7 Exercises on Local Convergence

2.7.1. Apply Newton’s method with (a) analytic first and second derivatives, (b) analytic first
derivatives and forward difference second derivatives, and (c) forward difference first and
second derivatives to find a local minimum of
1. f (x) = sin2 (x),

2
2. f (x) = ex , and

eta = .1 eta = .1
5 3
10 10
Function Value
Gradient Norm
0 2
10 10
−5 1
10 10
−10 0
10 10
0 5 10 0 5 10
eta = .0001 eta = .0001

5 3
10 10
0
10
Function Value
Gradient Norm
2
10
−5
10
1
10
−10
10
−15 0
10 10
0 2 4 6 8 0 2 4 6 8
Figure 2.3: Newton–CG for the Discrete Control Problem: η = .1, .0001
3. f (x) = x4 .
Use difference steps of h = 10−1 , 10−2 , 10−4 , and 10−8 . Explain your results.
2.7.2. Repeat part (c) of exercise 2.7.1. Experiment with
2
f (x) = ex + 10−4 rand(x) and f (x) = x2 + 10−4 rand(x),
where rand denotes the random number generator in your computing environment. Ex-
plain the differences in the results.
2.7.3. Show that if A is symmetric, p = 0, and pT Ap = 0, then A is either singular or indefinite.
2.7.4. Show that if b ∈ RN and the N × N matrix A is symmetric and has a negative eigenvalue,
then the quadratic functional
m(x) = xT Ax + xT b
does not have a minimizer. Show that if u is an eigenvector corresponding to a negative

eigenvalue of the Hessian, then u is a direction of negative curvature.
2.7.5. If N = 1, the local quadratic model could easily be replaced by a local quartic (i.e.,
fourth-degree) model (what would be wrong with a cubic model?). If a method is based
on minimization of the local quartic model, what kind of local convergence would you

expect? How would you extend this method to the case N > 1? Look at [30] for some
results on this.
2.7.6. Show that if the standard assumptions hold, h is sufficiently small, and x is sufficiently
near x∗ , the difference Hessian defined by (2.19), ∇2h f (x), is spd.
2.7.7. Write a locally convergent Newton method code based on accurate function and gradient
information and forward difference Hessians using (2.19). Be sure that your code tests for
positivity of the Hessian so that you can avoid convergence to a local maximum. Is the
test for positivity expensive? Apply your code to the parameter ID problem from §1.6.2.
If you use an ODE solver that lets you control the accuracy of the integration, try values
of the accuracy from 10−8 to 10−2 and see how the iteration performs. Be sure that your
difference Hessian reflects the accuracy in the gradient.
2.7.8. Let the standard assumptions hold and let λs > 0 be the smallest eigenvalue of ∇2 f (x∗ ).
Give the best (i.e., largest) bound you can for ρ such that ∇2 f (x) is positive definite for
all x ∈ B(ρ).
2.7.9. Use the definition of A† to prove (2.41).
2.7.10. Fill in the missing details in the proof of Theorem 2.4.2 by showing how the Kantorovich
theorem can be used to prove the existence of ξ(x).
2.7.11. Let f (x) = x2 and f (x) = sin(100x)/10. Using an initial iterate of x0 = 1, try to find
a local minimum of f + f using Newton’s method with analytic gradients and Hessians.
Repeat the experiment with difference gradients and Hessians (try forward differences
with a step size of h = .2).
2.7.12. Solve the parameter ID problem from §2.6 with the observations perturbed randomly (for
example, you could use the MATLAB rand function for this). Vary the amplitude of the
perturbation and see how the performance of Newton and Gauss–Newton changes.
2.7.13. Derive sensitivity equations for the entries of the Hessian for the parameter ID objective
function. In general, if there are P parameters, how many sensitivity equations would
need to be solved for the gradient? How many for the Hessian?
2.7.14. Solve the discrete control problem from §2.6.2 using Newton–CG with forcing terms that
depend on n. Consider ηn = .5/n, ηn = min(.1,
∇f (un )
), and some of the choices
from [99]. Vary N and the termination criteria and compare the performance with the
constant η choice in §2.6.2.
2.7.15. Let F : RN → RM (where M and N need not be the same) be sufficiently smooth (how
smooth is that?) and be such that F can also be computed for complex arguments. Show
that [181], [245]
Im(F (x + ihu))/h = F (x)u + O(h2 ),
where Im denotes imaginary part. What happens if there is error in F ? How can you use
this fact to compute better difference gradients and Hessians?


Chapter 3
Global Convergence
The locally convergent algorithms discussed in Chapter 2 can and do fail when the initial iterate
is not near the root. The reasons for this failure, as we explain below, are that the Newton
direction may fail to be a direction of descent for f and that even when a search direction is a
direction of decrease of f , as −∇f is, the length of the step can be too long. Hence, taking a
Newton (or Gauss–Newton, or inexact Newton) step can lead to an increase in the function and
divergence of the iteration (see exercise 3.5.14 for two dramatic examples of this). The globally
convergent algorithms developed in this chapter partially address this problem by either finding
a local minimum or failing in one of a small number of easily detectable ways.
These are not algorithms for global optimization. When these algorithms are applied to
problems with many local minima, the results of the iteration may depend in complex ways on
the initial iterate.
3.1 The Method of Steepest Descent

The steepest descent direction from x is d = −∇f (x). The method of steepest descent [52]
updates the current iteration xc by the formula
(3.1) x+ = xc − λ∇f (xc ).
If we take the simple choice λ = 1, then x+ is not guaranteed to be nearer a solution than xc ,
even if xc is very near a solution that satisfies the standard assumptions. The reason for this is
that, unlike the Newton direction, the steepest descent direction scales with f . The Newton step,
on the other hand, is the same for f as it is for cf for any c = 0 but need not be a direction of
decrease for f .
To make the method of steepest descent succeed, it is important to choose the steplength λ.
One way to do this, which we analyze in §3.2, is to let λ = β m , where β ∈ (0, 1) and m ≥ 0 is
the smallest nonnegative integer such that there is sufficient decrease in f . In the context of the
steepest descent algorithm, this means that
(3.2) f (xc − λ∇f (xc )) − f (xc ) < −αλ

∇f (xc )
2 .
This strategy, introduced in [7] and called the Armijo rule, is an example of a line search in
which one searches on a ray from xc in a direction in which f is locally decreasing. In (3.2),
α ∈ (0, 1) is a parameter, which we discuss after we fully specify the algorithm. This strategy
of repeatedly testing for sufficient decrease and reducing the stepsize if the test is failed is called
backtracking for obvious reasons.
39
The motivation for (3.2) is that if we approximate f by the linear model
mc = f (xc ) + ∇f (xc )(x − xc ),
then the reduction in the model (i.e., the predicted reduction in f ) is
pred = mc (xc ) − mc (x+ ) = λ

∇f (xc )
2 .
(3.2) says that the actual reduction in f
ared = f (xc ) − f (x+ )
is at least as much as a fraction of the predicted reduction in the linear model. The parameter α
is typically set to 10−4 .
The reason we demand sufficient decrease instead of simple decrease (i.e., f (xc ) < f (x+ )
or α = 0) is largely theoretical; a nonzero value of α is required within the proof to insure that
the iteration does not stagnate before convergence.
Algorithm 3.1.1. steep(x, f, kmax)
1. For k = 1, . . . , kmax
(a) Compute f and ∇f ; test for termination.
(b) Find the least integer m ≥ 0 such that (3.2) holds for λ = β m .
(c) x = x + λd.
2. If k = kmax and the termination test is failed, signal failure.
The termination criterion could be based on (2.12), for example.
3.2 Line Search Methods and the Armijo Rule

We introduce a few new concepts so that our proof of convergence of Algorithm steep will
also apply to a significantly more general class of algorithms.
Definition 3.2.1. A vector d ∈ RN is a descent direction for f at x if

df (x + td)
= ∇f (x)T d < 0.
dt t=0
Clearly the steepest descent direction d = −∇f (x) is a descent direction. A line search
algorithm searches for decrease in f in a descent direction, using the Armijo rule for stepsize
control, unless ∇f (x) = 0.
We will consider descent directions based on quadratic models of f of the form
1
m(x) = f (xc ) + ∇f (xc )T (x − xc ) + (x − xc )T Hc (x − xc ),
2
where Hc , which is sometimes called the model Hessian, is spd. We let d = x − xc be such that
m(x) is minimized. Hence,
∇m(x) = ∇f (xc ) + Hc (x − xc ) = 0
and hence
(3.3) d = −Hc−1 ∇f (xc ).

GLOBAL CONVERGENCE 41
The steepest descent direction satisfies (3.3) with Hc = I. However, the Newton direction
d = −∇2 f (x)−1 ∇f (x) may fail to be a descent direction if x is far from a minimizer because
∇2 f may not be spd. Hence, unlike the case for nonlinear equations [154], Newton’s method is
not a generally good global method, even with a line search, and must be modified (see [113],
[117], [231], and [100]) to make sure that the model Hessians are spd.
The algorithm we analyze in this section is an extension of Algorithm steep that allows for
descent directions that satisfy (3.3) for spd H. We modify (3.2) to account for H and the new
descent direction d = −H −1 ∇f (x). The general sufficient decrease condition is
(3.4) f (xc + λd) − f (xc ) < αλ∇f (xc )T d.
Here, as in (3.2), α ∈ (0, 1) is an algorithmic parameter. Typically α = 10−4 .

The stepsize reduction scheme in step 1b of Algorithm steep is crude. If β is too large, too
many stepsize reductions may be needed before a step is accepted. If β is too small, the progress
of the entire iteration may be retarded. We will address this problem in two ways. In §3.2.1 we
will construct polynomial models of f along the descent direction to predict an optimal factor
by which to reduce the step. In §3.3.3 we describe a method which remembers the steplength
from the previous iteration.
Our proofs require only the following general line search strategy. If a steplength λc has
been rejected (i.e., (3.4) fails with λ = λc ), construct
(3.5) λ+ ∈ [βlow λc , βhigh λc ],
where 0 < βlow ≤ βhigh < 1. The choice β = βlow = βhigh is the simple rule in Algo-
rithm steep. An exact line search, in which λ is the exact minimum of f (xc + λd), is not only
not worth the extra expense but can degrade the performance of the algorithm.
Algorithm 3.2.1. optarm(x, f, kmax)
1. For k = 1, . . . , kmax
(b) Construct an spd matrix H and solve (3.3) to obtain a descent direction d.
(c) Beginning with λ = 1, repeatedly reduce λ using any strategy that satisfies (3.5)
until (3.4) holds.
(d) x = x + λd.
2. If k = kmax and the termination test is failed, signal failure.
In the remainder of this section we prove that if the sequence of model Hessians remains
uniformly bounded and positive definite and the sequence of function values {f (xk )} is bounded
from below, then any limit point of the sequence {xk } generated by Algorithm optarm con-
verges to a point x∗ that satisfies the necessary conditions for optimality. We follow that analysis
with a local convergence theory that is much less impressive than that for Newton’s method.
We begin our analysis with a simple estimate that follows directly from the spectral theorem
for spd matrices.
Lemma 3.2.1. Let H be spd with smallest and largest eigenvalues 0 < λs < λl . Then for
all z ∈ RN ,
λ−1 2
l
z
≤ z H
T −1
z ≤ λ−1
s
z
.
2
The first step in the analysis is to use Lemma 3.2.1 to obtain a lower bound for the steplength.

Lemma 3.2.2. Assume that ∇f is Lipschitz continuous with Lipschitz constant L. Let
α ∈ (0, 1), x ∈ RN , and H be an spd matrix. Let λs > 0 be the smallest and λl ≥ λs the
largest eigenvalues of H. Let d be given by (3.3). Assume that ∇f (x) = 0. Then (3.4) holds for
any λ such that
2λs (1 − α)
(3.6) 0<λ≤ .
Lκ(H)
Proof. Let d = −H −1 ∇f (x). By the fundamental theorem of calculus

1
f (x + λd) − f (x) = ∇f (x + tλd)T λd dt.
0
Hence
f (x + λd) = f (x) + λ∇f (x)T d
(3.7) 1
+λ 0
(∇f (x + tλd) − ∇f (x))T d dt.
Therefore,
λ2 L
f (x + λd) = f (x − λH −1 ∇f (x)) ≤ f (x) + λ∇f (x)T d +
d
2 .
2
Positivity of H, Lemma 3.2.1, and κ(H) = λl λ−1
s imply that

d
2 =
H −1 ∇f (x)
2 ≤ λ−2 T
s ∇f (x) ∇f (x)
≤ −λl λ−2 T −1 T
s ∇f (x) d = −κ(H)λs ∇f (x) d.
Hence
f (x + λd) ≤ f (x) + (λ − λ2 Lλ−1 T
s κ(H)/2)∇f (x) d,
which implies (3.4) if

α ≤ (1 − λLλ−1
s κ(H)/2).
This is equivalent to (3.6).

Lemma 3.2.3. Let ∇f be Lipschitz continuous with Lipschitz constant L. Let {xk } be the
iteration given by Algorithm optarm with spd matrices Hk that satisfy
(3.8) κ(Hk ) ≤ κ̄
for all k. Then the steps
sk = xk+1 − xk = λk dk = −λk Hk−1 ∇f (xk )
satisfy
2βlow λs (1 − α)
(3.9) λk ≥ λ̄ =
Lκ̄
and at most
2λs (1 − α)
(3.10) m = log / log(βhigh )
Lκ̄
stepsize reductions will be required.
Proof. In the context of Algorithm optarm, Lemma 3.2.2 implies that the line search will
terminate when
2λs (1 − α)
λ≤ ,
Lκ(Hk )

if not before. The most that one can overshoot this is by a factor of βlow , which proves (3.9).
The line search will require at most m stepsize reductions, where m is the least nonnegative
integer such that
2λs (1 − α) m
> βhigh .
Lκ(Hk )
This implies (3.10).
The convergence theorem for Algorithm optarm says that if the condition numbers of the
matrices H and the norms of the iterates remain bounded, then every limit point of the iteration
is a stationary point. Boundedness of the sequence of iterates implies that there will be limit
points, but there is no guarantee that there is a unique limit point.
Theorem 3.2.4. Let ∇f be Lipschitz continuous with Lipschitz constant L. Assume that
the matrices Hk are spd and that there are κ̄ and λl such that κ(Hk ) ≤ κ̄, and
Hk
≤ λl for
all k. Then either f (xk ) is unbounded from below or
(3.11) lim ∇f (xk ) = 0
k→∞
and hence any limit point of the sequence of iterates produced by Algorithm optarm is a
stationary point.
In particular, if f (xk ) is bounded from below and xkl → x∗ is any convergent subsequence
of {xk }, then ∇f (x∗ ) = 0.
Proof. By construction, f (xk ) is a decreasing sequence. Therefore, if f (xk ) is bounded
from below, then limk→∞ f (xk ) = f ∗ exists and
(3.12) lim f (xk+1 ) − f (xk ) = 0.
k→∞
By (3.4) and Lemma 3.2.3 we have

f (xk+1 ) − f (xk ) < −αλk ∇f (xk )T Hk−1 ∇f (xk )
≤ −αλ̄λ−1 2
l
∇f (xk )
≤ 0.
Hence, by (3.12)
λl (f (xk ) − f (xk+1 ))

∇f (xk )
2 ≤ →0
αλ̄
as k → ∞. This completes the proof.
The analysis of the Armijo rule is valid for other line search methods [84], [125], [272],
[273]. The key points are that the sufficient decrease condition can be satisfied in finitely many
steps and that the stepsizes are bounded away from zero.
3.2.1 Stepsize Control with Polynomial Models

Having computed a descent direction d from xc , one must decide on a stepsize reduction scheme
for iterations in which (3.4) fails for λ = 1. A common approach [73], [84], [114], [197], [117]
is to model
ξ(λ) = f (xc + λd)
by a cubic polynomial. The data on hand initially are
ξ(0) = f (xc ), ξ (0) = ∇f (xc )T d < 0, and ξ(1) = f (x + d),
which is enough to form a quadratic model of ξ. So, if (3.4) does not hold with λ = λ0 = 1,
i.e.,
ξ(1) = f (xc + d) ≥ f (xc ) + α∇f (xc )T d = ξ(0) + αξ (0),

we approximate ξ by the quadratic polynomial
q(λ) = ξ(0) + ξ (0)λ + (ξ(1) − ξ(0) − ξ (0))λ2
and let λ1 be the minimum of q on the interval [βlow , βhigh ] ⊂ (0, 1). This minimum can be
computed directly since α ∈ (0, 1) and failure of (3.4) imply
q (λ) = 2(ξ(1) − ξ(0) − ξ (0)) > 2(α − 1)ξ (0) > 0.
Therefore, the global minimum of q is
−ξ (0)
λt = .
2(ξ(1) − ξ(0) − ξ (0))
So 

 βlow , λt ≤ βlow ,



(3.13) λ+ = λt , βlow < λt < βhigh ,





βhigh , λt ≥ βhigh .
If our first reduced value of λ does not satisfy (3.4), we base additional reductions on the
data
ξ(0) = f (xc ), ξ (0) = ∇f (xc )T d, ξ(λ− ), ξ(λc ),
where λc < λ− are the most recent values of λ to fail to satisfy (3.4). This is sufficient data to
approximate ξ with a cubic polynomial
q(λ) = ξ(0) + ξ (0)λ + c2 λ2 + c3 λ3 ,
where c2 and c3 can be determined by the equations
q(λc ) = ξ(λc ) = f (xc + λc d),
q(λ− ) = ξ(λ− ) = f (xc + λ− d),
which form the nonsingular linear system for c2 and c3

λ2c λ3c c2 ξ(λc ) − ξ(0) − ξ (0)λc
(3.14) = .
λ2− λ3− c3 ξ(λ− ) − ξ(0) − ξ (0)λ−
As with the quadratic case, q has a local minimum [84] at

−c2 + c22 − 3c3 ξ (0)
(3.15) λt = .
3c3
With λt in hand, we compute λ+ using (3.13). The application of (3.13) is called safeguarding
and is important for the theory, as one can see from the proof of Theorem 3.2.4. Safeguarding
is also important in practice because, if the cubic model is poor, the unsafeguarded model can
make steplength reductions that are so small that the iteration can stagnate or so large (i.e., near
1) that too many reductions are needed before (3.4) holds.

3.2.2 Slow Convergence of Steepest Descent

Unfortunately, methods based on steepest descent do not enjoy good local convergence prop-
erties, even for very simple functions. To illustrate this point we consider the special case of
convex quadratic objective functions
1 T
f (x) = x Ax − bT x + a,
2
where A is spd, b ∈ RN , and a is a scalar. We will look at a very simple example, using
the method of steepest descent with Hk = I (so λl = λs = 1) and show how the speed of
convergence depends on conditioning and scaling.
Lemma 3.2.5. Let f be a convex quadratic and let Hk = I for all k. Then the sequence
{xk } generated by Algorithm optarm converges to the unique minimizer of f .
Proof. In exercise 3.5.4 you are asked to show that f is bounded from below and that
∇f (x) = Ax − b. Hence ∇f (x∗ ) vanishes only at x∗ = A−1 b. Since ∇2 f (x) = A is spd, the
second-order sufficient conditions hold and x∗ is the unique minimizer of f .
Theorem 3.2.4 implies that
lim ∇f (xk ) = Axk − b = A(xk − x∗ ) = 0,

k→∞
and hence xk → x∗ .
Since the steepest descent iteration converges to the unique minimizer of a convex quadratic,
we can investigate the rate of convergence without concern about the initial iterate. We do this
in terms of the A-norm. The problems can be illustrated with the simplest case a = 0 and b = 0.
Proposition 3.2.6. Let f (x) = xT Ax/2 and let {xk } be given by Algorithm optarm with
Hk = I for all k. Then the sequence {xk } satisfies
(3.16)
xk+1
A = (1 − O(κ(A)−2 ))
xk
A .
Proof. The sufficient decrease condition, (3.4), implies that for all k
xTk+1 Axk+1 − xTk Axk = 2(f (xk+1 ) − f (xk ))
(3.17) ≤ 2α∇f (xk )T (xk+1 − xk )
= 2αλk ∇f (xk )T d = −2αλk (Axk )T (Axk ).
The Lipschitz constant of ∇f is simply λl =

A
; hence we may write (3.9) as
2β(1 − α)
(3.18) λk ≥ λ̄ = .
λl κ(A)
In terms of the A-norm, (3.17) can be written as

xk+1
2A −
xk
2A ≤ −2αλ̄λs
xk
2A ,
where we use the fact that

Az
2 = (Az)T (Az) ≥ λs z T Az = λs
z
2A .

Hence,

xk+1
2A ≤ (1 − 2αλ̄λs )
xk
2A ≤ (1 − 4α(1 − α)βκ(A)−2 )
xk
2A .
This completes the proof.

Now we consider two specific examples. Let N = 1 and define
f (x) = ωx2 /2,
where
(3.19) ω < 2(1 − α).
In this case x∗ = 0. We have ∇f (x) = f (x) = ωx and hence for all x ∈ R,
ωx2
f (x − ∇f (x)) − f (x) = ((1 − ω)2 − 1)
2
ω 2 x2
= (ω − 2)
2
< −α|f (x)|2 = −αω 2 x2
because (3.19) implies that

ω − 2 < −2α.
Hence (3.4) holds with d = ∇f (x) and λ = 1 for all x ∈ R. The rate of convergence can be
computed directly since
x+ = (1 − ω)xc
for all xc . The convergence is q-linear with q-factor 1 − ω. So if ω is very small, the convergence
will be extremely slow.
Similarly, if ω is large, we see that
ω 2 x2
f (x − λ∇f (x)) − f (x) = (λω − 2) < −αλω 2 x2
2
only if
2(1 − α)
λ< .
ω
So
2(1 − α) 2(1 − α)
β < βm = λ < .
ω ω
If ω is very large, many steplength reductions will be required with each iteration and the line
search will be very inefficient.
These are examples of poor scaling, where a change in f by a multiplicative factor can
dramatically improve the efficiency of the line search or the convergence speed. In fact, if
ω = 1, steepest descent and Newton’s method are the same and only one iteration is required.
The case for a general convex quadratic is similar. Let λl and λs be the largest and smallest
eigenvalues of the spd matrix A. We assume that b = 0 and a = 0 for this example. We let ul
and us be unit eigenvectors corresponding to the eigenvalues λl and λs . If
λs < 2(1 − α)
is small and the initial iterate is in the direction of us , convergence will require a very large
number of iterations (slow). If λl is large and the initial iterate is in the direction of ul , the line
search will be inefficient (many stepsize reductions at each iteration).

Newton’s method does not suffer from poor scaling of f and converges rapidly with no need
for a line search when the initial iterate is near the solution. However, when far away from the
solution, the Newton direction may not be a descent direction at all and the line search may
fail. Making the transition from steepest descent, which is a good algorithm when far from
the solution, to Newton’s or some other superlinearly convergent method as the iteration moves
toward the solution, is the central problem in the design of line search algorithms. The scaling
problems discussed above must also be addressed, even when far from the solution.
3.2.3 Damped Gauss–Newton Iteration

As we showed in §2.4, the steepest descent direction for the overdetermined least squares objec-
tive
M
1 1
f (x) =
ri (x)
22 = R(x)T R(x)
2 i=1 2
is
−∇f (x) = −R (x)T R(x).
The steepest descent algorithm could be applied to nonlinear least squares problems with the
good global performance and poor local convergence that we expect.
The Gauss–Newton direction at x
dGS = −(R (x)T R (x))−1 R (x)T R(x)
is not defined if R fails to have full column rank. If R does have full column rank, then
(dGS )T ∇f (x) = −(R (x)T R(x))T (R (x)T R (x))−1 R (x)T R(x) < 0,
and the Gauss–Newton direction is a descent direction. The combination of the Armijo rule with
the Gauss–Newton direction is called damped Gauss–Newton iteration.
A problem with the damped Gauss–Newton algorithm is that, in order for Theorem 3.2.4 to
be applicable, the matrices {R (xk )T R (xk )} must not only have full column rank but also must
be uniformly bounded and well conditioned, which are very strong assumptions (but if they are
satisfied, damped Gauss–Newton is a very effective algorithm).
The Levenberg–Marquardt method [172], [183] addresses these issues by adding a regular-
ization parameter ν > 0 to R (xc )T R (xc ) to obtain x+ = xc + s where
(3.20) s = −(νc I + R (xc )T R (xc ))−1 R (xc )T R(xc ),
where I is the N × N identity matrix. The matrix νc I + R (xc )T R (xc ) is positive definite.
The parameter ν is called the Levenberg–Marquardt parameter.
It is not necessary to compute R (xc )T R (xc ) to compute a Levenberg–Marquardt step [76].
One can also solve the full-rank (M + N ) × N linear least squares problem
2
1 R (xc ) R(xc )
(3.21) min √ s +
2 νc I 0
to compute s (see exercise 3.5.6). Compare this with computing the undamped Gauss–Newton
step by solving (2.33).
If one couples the Levenberg–Marquardt method with the Armijo rule, then Theorem 3.2.4
is applicable far from a minimizer and Theorem 2.4.1 nearby. We ask the reader to provide the
details of the proof in exercise 3.5.7.

Theorem 3.2.7. Let R be Lipschitz continuous. Let xk be the Levenberg–Marquardt–Armijo

iterates. Assume that
R (xk )
is uniformly bounded and that the sequence of Levenberg–
Marquardt parameters {νk } is such that
κ(νk I + R (xk )T R (xk ))
is bounded. Then
lim R (xk )T R(xk ) = 0.
k→∞
Moreover, if x∗ is any limit point of {xk } at which R(x∗ ) = 0, Assumption 2.4.1 holds, and
νk → 0, then xk → x∗ q-superlinearly. If, moreover,
νk = O(
R(xk )
)
as k → ∞ then the convergence is q-quadratic.

For example, if κ(R (xk )T R (xk )) and
R (xk )
are bounded then νk = min(1,
R(xk )
)
would satisfy the assumptions of Theorem 3.2.7. For a zero residual problem, this addresses the
potential conditioning problems of the damped Gauss–Newton method and still gives quadratic
convergence in the terminal phase of the iteration. The Levenberg–Marquardt–Armijo iteration
will also converge, albeit slowly, for a large residual problem.
We will not discuss globally convergent methods for underdetermined least squares problems
in this book. We refer the reader to [24], [252], and [253] for discussion of underdetermined
problems.
3.2.4 Nonlinear Conjugate Gradient Methods

Operationally, the conjugate gradient iteration for a quadratic problem updates the current iter-
ation with a linear combination of the current residual r and a search direction p. The search
direction is itself a linear combination of previous residuals. Only r and p need be stored to
continue the iteration. The methods discussed in this section seek to continue this idea to more
nonlinear problems.
Nonlinear conjugate gradient algorithms have the significant advantage of low storage over
most of the other algorithms covered in this book, the method of steepest descent being the
exception. For problems so large that the Newton or quasi–Newton methods cannot be imple-
mented using the available storage, these methods are among the few options (see [177] and [5]
for examples).
Linear conjugate gradient seeks to minimize f (x) = xT Hx/2 − xT b. The residual r =
b − Hx is simply −∇f (x), leading to a natural extension to nonlinear problems in which
r0 = p0 = ∇f (x0 ) and, for k ≥ 1,
(3.22) rk = ∇f (xk ) and pk = rk + βk pk−1 .
The update of x
xk+1 = xk + αk pk
can be done with a simple analytic minimization in the quadratic case, but a line search will be
needed in the nonlinear case. The missing pieces, therefore, are the choice of βk , the way the
line search will be done, and convergence theory. Theory is needed, for example, to decide if
pk is a descent direction for all k.
The general form of the algorithm is very simple. The inputs are an initial iterate, which
will be overwritten by the solution, the function to be minimized, and a termination vector
τ = (τr , τa ) of relative and absolute residuals.
Algorithm 3.2.2. nlcg(x, f, τ )

1. r0 =
∇f (x)
; k = 0
2. Do while
∇f (x)
> τr r0 + τa
(a) If k = 0 then p = −∇f (x) else

p = −∇f (x) + βp
(b) x = x + αp
The two most common choices for β, both of which are equivalent to the linear CG formula
in the quadratic case, are the Fletcher–Reeves [106]

∇f (xk )
2
(3.23) βkF R =

∇f (xk−1 )
2
and Polak–Ribière [215], [216]
∇f (xk )T (∇f (xk ) − ∇f (xk−1 ))

(3.24) βk =

∇f (xk−1 )
2
formulas. The Fletcher–Reeves method has been observed to take long sequences of very small
steps and virtually stagnate [112], [207], [208], [226]. The Polak–Ribière formula performs
much better and is more commonly used but has a less satisfactory convergence theory.
The line search has more stringent requirements, at least for proofs of convergence, than
are satisfied by the Armijo method that we advocate for steepest descent. We require that the
steplength parameter satisfies the Wolfe conditions [272], [273]
(3.25) f (xk + αk pk ) ≤ f (xk ) + σα αk ∇f (xk )T pk
and
(3.26) ∇f (xk + αk pk )T pk ≥ σβ ∇f (xk )T pk ,
where 0 < σα < σβ < 1. The first of the Wolfe conditions (3.25) is the sufficient decrease
condition, (3.4), that all line search algorithms must satisfy. The second (3.26) is often, but not
always, implied by the Armijo backtracking scheme of alternating a test for sufficient decrease
and reduction of the steplength. One can design a line search method that will, under modest
assumptions, find a steplength satisfying the Wolfe conditions [104], [171], [193].
The convergence result [3] for the Fletcher–Reeves formula requires a bit more. The proof
that pk is descent direction requires the strong Wolfe conditions, which replace (3.26) by
(3.27) |∇f (xk + αk pk )T pk | ≤ −σβ ∇f (xk )T pk
and demand that 0 < σα < σβ < 1/2. The algorithm from [193], for example, will find a point
satisfying the strong Wolfe conditions.
Theorem 3.2.8. Assume that the set
N = {x | f (x) ≤ f (x0 )}
is bounded and that f is Lipschitz continuously differentiable in a neighborhood of N . Let

Algorithm nlcg be implemented with the Fletcher–Reeves formula and a line search that satisfies
the strong Wolfe conditions. Then
lim ∇f (xk ) = 0.

This result has been generalized to allow for any choice of βk such that |βk | ≤ βkF R [112].
A similar result for the Polak–Ribière method, but with more complex conditions on the line
search, has been proved in [134]. This complexity in the line search is probably necessary, as
there are examples where reasonable line searches lead to failure in the Polak–Ribière method,
[222]. One can also prove convergence if βkP R is replaced by max(βkP R , 0) [112].
There is continuing research on these methods and we point to [112], [134], [205], and [202]
as good sources.
3.3 Trust Region Methods

Trust region methods overcome the problems that line search methods encounter with non-spd
approximate Hessians. In particular, a Newton trust region strategy allows the use of complete
Hessian information even in regions where the Hessian has negative curvature. The specific trust
region methods we will present effect a smooth transition from the steepest descent direction to
the Newton direction in a way that gives the global convergence properties of steepest descent
and the fast local convergence of Newton’s method.
The idea is very simple. We let ∆ be the radius of the ball about xc in which the quadratic
model
mc (x) = f (xc ) + ∇f (xc )T (x − xc ) + (x − xc )T Hc (x − xc )/2
can be trusted to accurately represent the function. ∆ is called the trust region radius and the
ball
T (∆) = {x |
x − xc
≤ ∆}
is called the trust region.
We compute the new point x+ by (approximately) minimizing mc over T (∆). The trust
region problem for doing that is usually posed in terms of the difference st between xc and the
minimizer of mc in the trust region
(3.28) min mc (xc + s).
s≤∆
We will refer to either the trial step st or the trial solution xt = xc + st as the solution to the
trust region problem.
Having solved the trust region problem, one must decide whether to accept the step and/or to
change the trust region radius. The trust region methods that we will discuss in detail approximate
the solution of the trust region problem with the minimizer of the quadratic model along a
piecewise linear path contained in the trust region. Before discussing these specific methods,
we present a special case of a result from [223] on global convergence.
A prototype trust region algorithm, upon which we base the specific instances that follow, is
Algorithm 3.3.1.
Algorithm 3.3.1. trbasic(x, f )
1. Initialize the trust region radius ∆.
2. Do until termination criteria are satisfied
(a) Approximately solve the trust region problem to obtain xt .
(b) Test both the trial point and the trust region radius and decide whether or not to
accept the step, the trust region radius, or both. At least one of x or ∆ will change
in this phase.
Most trust region algorithms differ only in how step 2a in Algorithm trbasic is done.
There are also different ways to implement step 2b, but these differ only in minor details and the
approach we present next in §3.3.1 is very representative.

3.3.1 Changing the Trust Region and the Step

The trust region radius and the new point are usually tested simultaneously. While a notion of
sufficient decrease is important, the test is centered on how well the quadratic model approximates
the function inside the trust region. We measure this by comparing the actual reduction in f
ared = f (xc ) − f (xt )
with the predicted reduction, i.e., the decrease in the quadratic model
pred = mc (xc ) − mc (xt ) = −∇f (xc )T st − sTt Hc st /2.
pred > 0 for all the trust region algorithms we discuss in this book unless ∇f (xc ) = 0. We
will introduce three control parameters
µ0 ≤ µlow < µhigh ,
which are used to determine if the trial step should be rejected (ared/pred < µ0 ) and/or the trust
region radius should be decreased (ared/pred < µlow ), increased (ared/pred > µhigh ), or left
unchanged. Typical values are .25 for µlow and .75 for µhigh . Both µ0 = 10−4 or µ0 = µlow
are used. One can also use the sufficient decrease condition (3.4) to determine if the trial step
should be accepted [84].
We will contract and expand the trust region radius by simply multiplying ∆ by constants
0 < ωdown < 1 < ωup .
Typical values are ωdown = 1/2 and ωup = 2. There are many other ways to implement a trust
region adjustment algorithm that also give global convergence. For example, the relative error
|pred − ared|/
∇f
can be used [84] rather than the ratio ared/pred. Finally we limit the
number of times the trust region radius can be expanded by requiring
(3.29) ∆ ≤ CT
∇f (xc )
,
for some CT > 1, which may depend on xc . This only serves to eliminate the possibility of
infinite expansion and is used in the proofs. Many of the dogleg methods which we consider
later automatically impose (3.29).
The possibility of expansion is important for efficiency in the case of poor scaling of f .
The convergence theory presented here [162] also uses the expansion phase in the proof of
convergence, but that is not essential. We will present the algorithm to test the trust region in a
manner, somewhat different from much of the literature, that only returns once a new iterate has
been accepted.
Algorithm 3.3.2. trtest(xc , xt , x+ , f, ∆)
1. z = xc
2. Do while z = xc
(a) ared = f (xc ) − f (xt ), st = xt − xc , pred = −∇f (xc )T st − sTt Hc st /2
(b) If ared/pred < µ0 then set z = xc , ∆ = ωdown ∆, and solve the trust region
problem with the new radius to obtain a new trial point. If the trust region radius
was just expanded, set z = xold
t .
(c) If µ0 ≤ ared/pred < µlow , then set z = xt and ∆ = ωdown ∆.
(d) If µlow ≤ ared/pred ≤ µhigh , set z = xt .

(e) If µhigh < ared/pred and

st
= ∆ ≤ CT
∇f (xc )
, then set z = xc , ∆ = ωup ∆,
and solve the trust region problem with the new radius to obtain a new trial point.
Store the old trial point as xold
t in case the expansion fails.
3. x+ = z.
The loop in Algorithm trtest serves the same purpose as the loop in a line search algorithm
such as Algorithm steep. One must design the solution to the trust region problem in such a
way that that loop will terminate after finitely many iterations and a general way to do that is the
subject of the next section.
We incorporate Algorithm trtest into a general trust region algorithm paradigm that we
will use for the remainder of this section.
Algorithm 3.3.3. trgen(x, f )
1. Initialize ∆
2. Do forever
(a) Let xc = x. Compute ∇f (xc ) and an approximate Hessian Hc .
(b) Solve the trust region problem to obtain a trial point xt .
(c) Call trtest(xc , xt , x, f, ∆)
Hessians and gradients are computed only in step 2a of Algorithm trgen.
3.3.2 Global Convergence of Trust Region Algorithms

While one can, in principal, solve the trust region problem exactly (see §3.3.4), it is simpler
and more efficient to solve the problem approximately. It is amazing that one need not do a
very good job with the trust region problem in order to get global (and even locally superlinear)
convergence.
Our demands of our solutions of the trust region problem and our local quadratic models
are modest and readily verifiable. The parameter σ in part 1 of Assumption 3.3.1, like the
parameter CT in (3.29), is used in the analysis but plays no role in implementation. In the
specific algorithms that we discuss in this book, σ can be computed. Part 2 follows from well-
conditioned and bounded model Hessians if Algorithm trtest is used to manage the trust
region.
Assumption 3.3.1.
1. There is σ > 0 such that
(3.30) pred = f (xc ) − mc (xt ) ≥ σ

∇f (xc )
min(
st
,
∇f (xc )
).
2. There is M > 0 such that either

st
≥
∇f (xc )
/M or
st
= ∆c .
The global convergence theorem based on this assumption should be compared with the
similar result on line search methods—Theorem 3.2.4.
Theorem 3.3.1. Let ∇f be Lipschitz continuous with Lipschitz constant L. Let {xk }
be generated by Algorithm trgen and let the solutions for the trust region problems satisfy
Assumption 3.3.1. Assume that the matrices {Hk } are bounded. Then either f is unbounded
from below, ∇f (xk ) = 0 for some finite k, or
(3.31) lim ∇f (xk ) = 0.

k→∞

Proof. Assume that ∇f (xk ) = 0 for all k and that f is bounded from below. We will show
that there is MT ∈ (0, 1] such that once an iterate is taken (i.e., the step is accepted and the trust
region radius is no longer a candidate for expansion), then
(3.32)
sk
≥ MT
∇f (xk )
.
Assume (3.32) for the present. Since sk is an acceptable step, Algorithm trtest and part 1 of
Assumption 3.3.1 imply that
aredk ≥ µ0 predk ≥ µ0
∇f (xk )
σ min(
sk
,
∇f (xk )
).
We may then use (3.32) to obtain
(3.33) aredk ≥ µ0 σMT

∇f (xk )
2 .
Now since f (xk ) is a decreasing sequence and f is bounded from below, limk→∞ aredk = 0.
Hence (3.33) implies (3.31).
It remains to prove (3.32). To begin note that if
sk
< ∆k then by part 2 of Assumption 3.3.1

sk
≥
∇f (xk )
/M.
Hence, we need only consider the case in which
(3.34)
sk
= ∆k and
sk
<
∇f (xk )
,
since if (3.34) does not hold then (3.32) holds with MT = min(1, 1/M ).
We will complete the proof by showing that if (3.34) holds and sk is accepted, then
−2
2σ min(1 − µhigh , (1 − µ0 )ωup )
(3.35)
sk
= ∆k ≥
∇f (xk )
.
M +L
This will complete the proof with
−2
2σ min(1 − µhigh , (1 − µ0 )ωup )
MT = min 1, 1/M, .
M +L
Now increase the constant M > 0 in part 1 of Assumption 3.3.1 if needed so that
(3.36)
Hk
≤ M for all k.
We prove (3.35) by showing that if (3.34) holds and (3.35) does not hold for a trial step st ,
then the trust region radius will be expanded and the step corresponding to the larger radius will
be acceptable. Let st be a trial step such that
st
<
∇f (xk )
and
−2
2σ min(1 − µhigh , (1 − µ0 )ωup )
(3.37)
st
= ∆t <
∇f (xk )
.
M +L
We use the Lipschitz continuity of ∇f and (3.36) to obtain
1
aredt = −∇f (xk )T st − (∇f (xk + tst ) − ∇f (xk ))T st dt
0
1
= predt + sTt Hk st /2 − (∇f (xk + tst ) − ∇f (xk ))T st dt
0
≥ predt − (M + L)
st
2 /2.

Therefore, using (3.30) from Assumption 3.3.1, we have
aredt (M + L)
st
2
≥1−
predt 2predt
(3.38)
(M + L)
st
2
≥1− .
2σ
∇f (xk )
min(
∇f (xk )
,
st
)
Now since
st
<
∇f (xk )
by (3.34) we have
min(
∇f (xk )
,
st
) =
st
and hence
aredk (M + L)
st
(3.39) ≥1− > µhigh

predk 2
∇f (xk )
σ
by (3.37). Hence, an expansion step will be taken by replacing ∆t by ∆+
t = ωup ∆t and st by
s+
t , the minimum of the quadratic model in the new trust region.
Now (3.38) still holds and, after the expansion,

s+
t
≤ ωup
st
< ωup
∇f (xk )
.
So
min(
∇f (xk )
,
s+ +
t
) >
st
/ωup .
Hence,
ared+
t (M + L)
s+t
2
≥1−
pred+
t 2σ
∇f (xk )
min(
∇f (xk )
,
s+
t
)
2
(M + L)ωup
s+t
(M + L)ωup
st
≥1− ≥1− ≥ µ0
2
∇f (xk )
σ 2
∇f (xk )
σ
by (3.37). Hence, the expansion will produce an acceptable step. This means that if the final
accepted step satisfies (3.34), it must also satisfy (3.35). This completes the proof.
3.3.3 A Unidirectional Trust Region Algorithm

The most direct way to compute a trial point that satisfies Assumption 3.3.1 is to mimic the line
search and simply minimize the quadratic model in the steepest descent direction subject to the
trust region bound constraints.
In this algorithm, given a current point xc and trust region radius ∆c , our trial point is the
minimizer of
ψc (λ) = mc (xc − λ∇f (xc ))
subject to the constraint that
x(λ) = xc − λ∇f (xc ) ∈ T (∆c ).
Clearly the solution is x(λ̂) where

 ∆c

 ∇f (xc ) if ∇f (xc )T Hc ∇f (xc ) ≤ 0 ,
(3.40) λ̂ =

 min ∇f (xc )2 ∆c
T ,
∇f (xc ) Hc ∇f (xc ) ∇f (xc )
if ∇f (xc )T Hc ∇f (xc ) > 0.

x(λ̂), the minimizer of the quadratic model in the steepest descent direction, subject to the trust
region bounds, is called the Cauchy point. We will denote the Cauchy point by xCP c .
1
CP
Then with x as trial point, one can use Theorem 3.3.1 to derive a global convergence
theorem for the unidirectional trust region.
Theorem 3.3.2. Let ∇f be Lipschitz continuous with Lipschitz constant L. Let {xk } be
generated by Algorithm trgen with xt = xCP and (3.40). Assume that the matrices {Hk } are
bounded. Then either f (xk ) is unbounded from below, ∇f (xk ) = 0 for some finite k, or
lim ∇f (xk ) = 0.
k→∞
Proof. We show that xt satisfies part 2 of Assumption 3.3.1. If

st
= ∆c then the assertion
holds trivially. If
st
< ∆c then, by definition of xCP
c ,

∇f (xc )
2 ∇f (xc )
st = − .
∇f (xc )T Hc ∇f (xc )
Hence, if
Hc
≤ M ,

st
≥
∇f (xc )
/M
as asserted.
We leave the proof that xt satisfies part 1 for the reader (exercise 3.5.8).
The assumptions we used are stronger that those in, for example, [104] and [223], where
lim inf
∇f (xk )
= 0
rather than ∇f (xk ) → 0 is proved.
3.3.4 The Exact Solution of the Trust Region Problem

The theory of constrained optimization [117], [104] leads to a characterization of the solutions
of the trust region problem. In this section we derive that characterization via an elementary
argument (see also [84], [242], and [109]). This book focuses on approximate solutions, but the
reader should be aware that the exact solution can be computed accurately [192], [243].
Theorem 3.3.3. Let g ∈ RN and let A be a symmetric N × N matrix. Let
m(s) = g T s + sT As/2.
A vector s is a solution to
(3.41) min m(s)
s≤∆
if and only if there is ν ≥ 0 such that
(A + νI)s = −g
and either ν = 0 or
s
= ∆.
Proof. If
s
< ∆ then ∇m(s) = g + As = 0, and the conclusion follows with ν = 0. To
consider the case where
s
= ∆, let λ1 ≤ λ2 ≤ · · · λN be the eigenvalues of A.
1 In some of the literature, [84], for example, H is assumed to be positive definite and the Cauchy point is taken to
c
be the global minimizer of the quadratic model.

Clearly, for any ν,
m(s) = g T s + sT As/2
= g T s + sT (A + νI)s/2 − ν∆2 /2.
Consider the function, defined for ν > ν0 = max(0, −λ1 ),
s(ν) = −(A + νI)−1 g.
Since
lim s(ν) = 0
ν→∞
and
s(ν)
is a continuous decreasing function of ν ∈ (ν0 , ∞) we see that if
lim
s(ν)
> ∆
ν→ν0
then there is a unique ν such that

s(ν)
= ∆. Since ν ≥ ν0 , A + νI is positive semidefinite;
therefore, s(ν) is a global minimizer of
g T s + sT (A + νI)s/2.
Hence, we must have

m(s) ≥ m(s(ν))
for all s such that
s
= ∆. Hence, s(ν) is a solution of (3.41).
The remaining case is
lim
s(ν)
≤ ∆.
ν→ν0
This implies that g is orthogonal to the nontrivial space S0 of eigenfunctions corresponding to

−ν0 (for otherwise the limit would be infinite). If we let s = s1 + s2 , where s2 is the projection
of s onto S0 , we have
m(s) = sT1 g + sT1 (A + ν0 )s1 /2 + sT2 (A + ν0 )s2 /2 − ν0 ∆2 /2
= sT1 g + sT1 (A + λ0 )s1 /2 − ν0 ∆2 /2.
Hence, m(s) is minimized by setting s1 equal to the minimum norm solution of (A+ν0 )x = −g
(which exists by orthogonality of g to S0 ) and letting s2 be any element of S0 such that

s2
2 = ∆2 −
s1
2 .
This completes the proof.
3.3.5 The Levenberg–Marquardt Parameter

The solution of the trust region problem presented in §3.3.4 suggests that, rather than controlling
∆, one could set
st = −(νI + Hc )−1 g,
adjust ν in response to ared/pred instead of ∆, and still maintain global convergence. A natural
application of this idea is control of the Levenberg–Marquardt parameter. This results in a
much simpler algorithm than Levenberg–Marquardt–Armijo in that the stepsize control can be
eliminated. We need only vary the Levenberg–Marquardt parameter as the iteration progresses.

We present the algorithm from [190] to illustrate this point. For an inexact formulation, see
[276].
The Levenberg–Marquardt quadratic model of least squares objective
M
1 1
f (x) =
ri (x)
22 = R(x)T R(x)
2 i=1 2
with parameter νc at the point xc is
mc (x) = f (xc ) + (x − xc )T R (xc )T R(xc )

(3.42)
+ 12 (x − xc )T (R (xc )T R (xc ) + νc I)(x − xc ).
The minimizer of the quadratic model is the trial point
(3.43) xt = xc − (R (xc )T R (xc ) + νc I)−1 R (xc )T R(xc ),
the step is s = xt − xc , and the predicted reduction is
pred = m(xc ) − m(xt ) = −sT R (xc )T R(xc ) − 21 sT (R (xc )T R (xc ) + νc I)s
= −sT R (xc )T R(xc ) + 12 sT R (xc )T R(xc ) = − 12 sT ∇f (xc ).
The algorithm we present below follows the trust region paradigm and decides on accepting
the trial point and on adjustments in the Levenberg–Marquardt parameter by examinaing the
ratio
ared f (xc ) − f (xt )
=
pred m(xc ) − m(xt )
f (xc ) − f (xt )
= −2 .
sT ∇f (xc )
In addition to the trust region parameters 0 < ωdown < 1 < ωup and µ0 ≤ µlow < µhigh
we require a default value ν0 of the Levenberg–Marquardt parameter.
The algorithm for testing the trial point differs from Algorithm trtest in that we decrease
(increase) ν rather that increasing (decreasing) a trust region radius if ared/pred is large (small).
We also attempt to set the Levenberg–Marquardt parameter to zero when possible in order to
recover the Gauss–Newton iteration’s fast convergence for small residual problems.
Algorithm 3.3.4. trtestlm(xc , xt , x+ , f, ν)
1. z = xc
2. Do while z = xc
(a) ared = f (xc ) − f (xt ), st = xt − xc , pred = −∇f (xc )T st /2.

(b) If ared/pred < µ0 then set z = xc , ν = max(ωup ν, ν0 ), and recompute the trial
point with the new value of ν.
(c) If µ0 ≤ ared/pred < µlow , then set z = xt and ν = max(ωup ν, ν0 ).
(d) If µlow ≤ ared/pred, then set z = xt .
If µhigh < ared/pred, then set ν = ωdown ν.
If ν < ν0 , then set ν = 0.
3. x+ = z.

The Levenberg–Marquardt version ofAlgorithm trgen is simple to describe and implement.

Algorithm 3.3.5. levmar(x, R, kmax)
1. Set ν = ν0 .
2. For k = 1, . . . , kmax
(a) Let xc = x.
(b) Compute R, f , R , and ∇f ; test for termination.
(c) Compute xt using (3.43).
(d) Call trtestlm(xc , xt , x, f, ν)
We state a convergence result [190], [276] without proof.

Theorem 3.3.4. Let R be Lipschitz continuously differentiable. Let {xk } and {νk } be the
sequence of iterates and Levenberg–Marquardt parameters generated by Algorithm levmar
with kmax = ∞. Assume that {νk } is bounded from above. Then either R (xk )T R(xk ) = 0
for some finite k or
lim R (xk )T R(xk ) = 0.
k→∞
Moreover, if x is a limit point of {xk } for which R(x∗ ) = 0 and R (x∗ ) has full rank, then
∗
xk → x∗ q-quadratically and νk = 0 for k sufficiently large.
3.3.6 Superlinear Convergence: The Dogleg

The convergence of the unidirectional trust region iteration can be as slow as that for steepest
descent. To improve the convergence speed in the terminal phase we must allow for approx-
imations to the Newton direction. The power of trust region methods is the ease with which
the transition from steepest descent, with its good global properties, to Newton’s method can be
managed.
We define the Newton point at xc as
xN −1
c = xc − Hc ∇f (xc ).
If Hc is spd, the Newton point is the global minimizer of the local quadratic model. On the other
hand, if Hc has directions of negative curvature the local quadratic model will not have a finite
minimizer, but the Newton point is still useful. Note that if H = I the Newton point and the
Cauchy point are the same if the Newton point is inside the trust region.
We will restrict our attention to a special class of algorithms that approximate the solution
of the trust region problem by minimizing mc along a piecewise linear path S ⊂ T (∆). These
paths are sometimes called doglegs because of the shapes of the early examples [84], [80], [218],
[217], [220]. In the case where ∇2 f (x) is spd, one may think of the dogleg path as a piecewise
linear approximation to the path with parametric representation
{x − (λI + ∇2 f (x))−1 ∇f (x) | 0 ≤ λ}.
This is the path on which the exact solution of the trust region problem lies.
The next step up from the unidirectional path, the classical dogleg path [220], has as many
as three nodes, xc , xCP
c
∗
, and xN CP ∗
c . Here xc is the global minimizer of the quadratic model in
the steepest descent direction, which will exist if and only if ∇f (xc )T Hc ∇f (xc ) > 0. If xCP
c
∗
exists and
(3.44) (xNc − xc
CP ∗ T
) (xCPc
∗
− xc ) > 0,

we will let xN
c be the terminal node. If (3.44) holds, as it always will if Hc is spd, then the path
can be parameterized by the distance from xc and, moreover, mc decreases along the path. If
(3.44) does not hold, we do not use xN c as a node and revert to the unidirectional path in the
steepest descent direction.
Note that (3.44) implies
(3.45) ∇f (xc )T (xN c − xc ) < 0.
We can express the conditions for using the three node path rather than the unidirectional path
very simply. If xCP
c is on the boundary of the trust region then we accept xCP c as the trial point.
CP CP ∗
If xc = xc is in the interior of the trust region, then we test (3.44) to decide what to do.
With this in mind our trial point for the classical dogleg algorithm will be
 CP

 x if
xc − xCP
c
=∆

 CP ∗

 or x exists and (3.44) fails,



(3.46) xD (∆) = xN if
xc − xCP N
c
<
xc − xc
≤ ∆



 and (3.44) holds,




 D
y (∆) otherwise.
Here y D (∆) is the unique point between xCP

c and xN D
c such that
x − xc
= ∆.
The important properties of dogleg methods are as follows:
• No two points on the path have the same distance from xc ; hence the path may be param-
eterized as x(s), where s =
x(s) − xc
.
• mc (x(s)) is a strictly decreasing function of s.
This enables us to show that the dogleg approximate solution of the trust region problem sat-
isfies Assumption 3.3.1 and apply Theorem 3.3.1 to conclude global convergence. Superlinear
convergence will follow if Hk is a sufficiently good approximation to ∇2 f (xk ).
Lemma 3.3.5. Let xc , Hc , and ∆ be given. Let Hc be nonsingular,
sN = −Hc−1 ∇f (xc ), and xN = xc + sN .
Assume that ∇f (xc )T Hc ∇f (xc ) > 0 and let

∇f (xc )
2
sCP ∗ = xCP ∗ − xc = − ∇f (xc ).
Let S be the piecewise linear path from xc to xCP ∗ to xN . Then if
(3.47) (sN − sCP ∗ )T sCP ∗ > 0,
for any δ ≤
sN
there is a unique point x(δ) on S such that

x(δ) − xc
= δ.
Proof. Clearly the statement of the result holds on the segment of the path from x to xCP ∗ .
To prove the result on the segment from xCP ∗ to xN we must show that
1
φ(λ) =
(1 − λ)sCP ∗ + λsN
2
2

is strictly monotone increasing for λ ∈ (0, 1).

Since (3.47) implies that

sN
sCP ∗
≥ (sN )T sCP ∗ >
sCP ∗
2
and therefore that
sN
>
sCP ∗
, we have
φ (λ) = (sN − sCP ∗ )T ((1 − λ)sCP ∗ + λsN )
= −(1 − λ)
sCP ∗
2 + (1 − λ)(sN )T sCP ∗ + λ
sN
2 − λ(sN )T sCP ∗
> λ(
sN
2 − (sN )T sCP ∗ ) ≥ λ(
sN
−
sCP ∗
)
sN
> 0.
Hence, φ is an increasing function and the proof is complete.
The next stage is to show that the local quadratic model decreases on the dogleg path S.
Lemma 3.3.6. Let the assumptions of Lemma 3.3.5 hold. Then the local quadratic model
1
mc (x) = f (xc ) + ∇f (xc )T (x − xc ) + (x − xc )T Hc (x − xc )
2
is strictly monotone decreasing on S.
Proof. Since xCPc
∗
is the minimum of the local quadratic model in the steepest descent
direction, we need only show that mc is strictly decreasing on the segment of the path between
xCP
c
∗
and xN . Set
ψ(λ) = mc (xc + (1 − λ)sCP ∗ + λsN )
= f (xc ) + ∇f (xc )T ((1 − λ)sCP ∗ + λsN )
+ 12 ((1 − λ)sCP ∗ + λsN )T Hc ((1 − λ)sCP ∗ + λsN ).
Noting that Hc sN = −∇f (xc ) and sCP ∗ = −λ̂∇f (xc ), we obtain
ψ(λ) = f (xc ) − λ̂(1 − λ)2

∇f (xc )
2
+λ(1 − λ/2)∇f (xc )T sN
+ 12 (1 − λ)2 λ̂2 ∇f (xc )T Hc ∇f (xc ).

Therefore,
ψ (λ) = 2λ̂(1 − λ)
∇f (xc )
2
+(1 − λ)∇f (xc )T sN − (1 − λ)λ̂2 ∇f (xc )T Hc ∇f (cc ).

Since
λ̂∇f (xc )T Hc ∇f (cc ) =
∇f (xc )
2
we have, using (3.44),
ψ (λ) = (1 − λ)(λ̂
∇f (xc )
2 − ∇f (xc )T Hc−1 ∇f (xc ))
= (1 − λ)∇f (xc )T (λ̂∇f (xc ) − Hc−1 ∇f (xc ))
1−λ
= (xc − xCP
c
∗ T
) (xN
c − xc ) < 0,
λ̂

completing the proof.

At this point we have shown that the approximate trust region problem has a unique solution.
To prove global convergence we need only verify that the approximate solution of the trust region
problem xD satisfies Assumption 3.3.1.
generated by Algorithm trgen and the solutions for the trust region problem be given by
(3.46). Assume that the matrices {Hk } are bounded. Then either f (xk ) is unbounded from
below, ∇f (xk ) = 0 for some finite k, or
(3.48) lim ∇f (xk ) = 0.

k→∞
Proof. We need to check that the solutions of the trust region problem satisfy Assump-
tion 3.3.1. Part 2 of the assumption follows from the definition, (3.46), of xD and the bounded-
ness of the approximate Hessians. Let

Hk
≤ M
for all k. If
sk
< ∆, then (3.46) implies that (3.44) must hold and so xt = xN
k is the Newton
point. Hence,
−1

sk
=
xk − xN k
=
Hk ∇f (xk )
≥
∇f (xk )
/M,
which proves part 2.

Verification of part 1 will complete the proof. There are several cases to consider depending
on how xD is computed.
If xD = xCP then either
sCP
= ∆c or (3.44) fails. We first consider the case where
∇f (xc )T Hc ∇f (xc ) ≤ 0. In that case
sCP
= ∆c and λ̂ = ∆c /
∇f (xc )
. Therefore,
λ̂2
pred = λ̂
∇f (xc )
2 − T
2 ∇f (xc ) Hc ∇f (xc )
= ∆c
∇f (xc )
− ∆2c
2
∇f (xc )
2
≥ ∆c
∇f (xc )
=
s
∇f (xc )
.
Hence (3.30) holds with σ = 1.
Now assume that ∇f (xc )T Hc ∇f (xc ) > 0 and
sCP
= ∆c . In this case

∇f (xc )
2 ∆c
T
≥
∇f (xc ) Hc ∇f (xc )
∇f (xc )
and so
λ̂2
pred = λ̂
∇f (xc )
2 − T
2 ∇f (xc ) Hc ∇f (xc )
= ∆c
∇f (xc )
− ∆2c
2
∇f (xc )
2
≥ ∆c
∇f (xc )
/2,
which is (3.30) with σ = 1/2.
If (3.44) fails, ∇f (xc )T Hc ∇f (xc ) > 0, and
sCP
< ∆c , then

∇f (xc )
2
λ̂ = ,

and
λ̂2
pred = λ̂
∇f (xc )
2 − T
2 ∇f (xc ) Hc ∇f (xc )

∇f (xc )
4 λ̂
∇f (xc )
2
= =
2∇f (xc )T Hc ∇f (xc ) 2
∇f (xc )
= ,
2
which is (3.30) with σ = 1/2.
The final case is if (3.44) holds and xD = xCP . In that case the predicted reduction is more
than Cauchy decrease, i.e., the decrease obtained by taking the Cauchy point, and hence

∇f (xc )
4
pred ≥
2∇f (xc )T Hc ∇f (xc )

∇f (xc )
2
≥ ,
2M
which is (3.30) with σ = 1/(2M ). This completes the proof.
The last part of the proof of this theorem is very important, asserting that any solution of
the trust region problem for which pred is at least a fixed fraction of Cauchy decrease will give
global convergence. We refer the reader to [232] and [104] for a more general and detailed
treatment using this point of view.
Corollary 3.3.8. Any algorithm for solving the trust region problem that satisfies for some
τ >0
pred ≥ τ (mc (xc ) − mc (xCP c ))
satisfies (3.30) for σ = τ /2.

The trust region CG algorithm we present in §3.3.7 can be analyzed with this corollary.
If Hk = ∇2 f (xk ) or a sufficiently good approximation, then the classical dogleg will become
Newton’s method (or a superlinearly convergent method) as the iterations approach a minimizer
that satisfies the standard assumptions. Hence, the algorithm makes a smooth and automatic
transition into the superlinearly convergent stage.
generated by Algorithm trgen and the solutions for the trust region problem are given by (3.46).
Assume that Hk = ∇2 f (xk ) and that the matrices {Hk } are bounded. Let f be bounded from
below. Let x∗ be a minimizer of f at which the standard assumptions hold. Then if x∗ is a limit
point of xk , then xk → x∗ and the convergence is locally q-quadratic.
Proof. Since x∗ is a limit point of {xk }, there is, for any ρ > 0, a k sufficiently large so that

ek
< ρ,
Hk
≤ 2
∇2 f (x∗ )
,
Hk−1
≤ 2
(∇2 f (x∗ ))−1
,
and xk is near enough for the assumptions of Theorem 2.3.2 to hold. If Hk is spd, so is Hk−1
and for such k, (3.44) holds. Hence, the dogleg path has the nodes xk , xCP N
k , and xk . Moreover,
if ρ is sufficiently small, then

Hk−1 ∇f (xk )
≤ 2
ek
≤ 2ρ.
We complete the proof by showing that if ρ is sufficiently small, the trust region radius will be
expanded if necessary until the Newton step is in the trust region. Once we do this, the proof is
complete as then the local quadratic convergence of Newton’s method will take over.

Now
predk ≥
sk
∇f (xk )
/2
by the proof of Theorem 3.3.7. Using Hk = ∇2 f (xk ) we have
1
T
aredk = −∇f (xk ) stk − (∇f (xk + tstk ) − ∇f (xk ))T stk dt
0
1
= predk + sTtk ∇2 f (xk )stk /2 − (∇f (xk + tstk ) − ∇f (xk ))T stk dt
0
= predk + O(
sk
∇f (xk )
ρ)
and therefore ared/pred = 1−O(ρ). Hence, for ρ sufficiently small, the trust region radius will
be increased, if necessary, until the Newton point is inside the trust region and then a Newton
step will be taken. This completes the proof.
The classical dogleg algorithm is implemented in Algorithm ntrust, which uses the trust
radius adjustment scheme from Algorithm trtest. It is to be understood that trtest is
implemented so that xt is given by (3.46) and hence trtest only samples points on the
piecewise linear search path determined by the Cauchy point, the Newton point, and (3.44).
Algorithm 3.3.6. ntrust(x, f, τ )
1. Compute f (x) and ∇f (x)
2. τ = τa + τr
∇f (x)
3. Do while
∇f (x)
> τ
(a) Compute and factor ∇2 f (x)
(b) Compute the Cauchy and Newton points and test (3.44)
(c) Call trtest(x, xt , x+ , f, ∆)
(d) Compute f (x+ ) and ∇f (x+ ); x = x+
We implement Algorithm ntrust in the collection of MATLAB codes.
3.3.7 A Trust Region Method for Newton–CG

In this section we present a brief account of an algorithm from [247] (see also [257]) that combines
the trust region paradigm of §3.3.6 with the inexact Newton ideas of §2.5.2. We follow §2.5.2
and denote the preconditioner by M and let C = M −1 . We solve the scaled trust region problem
min φ(d),
dC ≤∆
where the quadratic model is still

1
φ(d) = ∇f (x)T d + dT ∇2 f (x)d.
2
Here the C-norm is

d
C = (dT Cd)1/2 .
The algorithmic description of the trust region problem solver from the TR–CG method
given below is from [162]. In [247] the algorithm is expressed in terms of C rather than M .

This is a dogleg method in that the approximate solution of the trust region problem lies on a
piecewise linear path with the CG iterations as nodes. As long as CG is performing properly
(i.e., pT w > 0) nodes are added to the path until the path intersects the trust region boundary. If
a direction of indefiniteness is found (pT w ≤ 0), then that direction is followed to the boundary.
In this way a negative curvature direction, if found in the course of the CG iteration, can be
exploited.
The inputs to Algorithm trcg are the current point x, the objective f , the forcing term η,
and the current trust region radius ∆. The output is the approximate solution of the trust region
problem d. This algorithm is not the whole story, as once the trust region problem is solved
approximately, one must use f (xc + d) to compute ared and then make a decision on how the
trust region radius should be changed. Our formulation differs from that in [247] in that the
termination criterion measures relative residuals in the l2 -norm rather than in the C-norm. This
change in the norm has no effect on the analysis in [247], and, therefore, we can apply the results
in §2.5 directly to draw conclusions about local convergence.
Algorithm 3.3.7. trcg(d, x, f, M, η, ∆, kmax)
1. r = −∇f (x), ρ0 =
r
22 , k = 1, d = 0
√
∇f (x)
2 and k < kmax
(a) z = M r
(b) τk−1 = z T r
else
β = τk−1 /τk−2 , p = z + βp
(d) w = ∇2 f (x)p
If pT w ≤ 0 then
Find τ such that
d + τ p
C = ∆
d = d + τ p; return
(e) α = τk−1 /pT w
(f) r = r − αw
(g) ρk = rT r
(h) dˆ = d + αp
ˆ C > ∆ then
(i) If
d
Find τ such that

d + τ p
C = ∆
d = d + τ p; return
(j) d = d;ˆ k =k+1
Algorithm trcg does what we would expect a dogleg algorithm to do in that the piecewise
linear path determined by the iteration moves monotonically away from x (in the
·
C -norm!)
and the quadratic model decreases on that path [247]. Algorithm trcg will, therefore, compute
the same Newton step as Algorithm fdpcg. One might think that it may be difficult to compute
the C-norm if one has, for example, a way to compute the action of M on a vector that does
not require computation of the matrix C. However, at the cost of storing two additional vectors
we can update Cp and Cd as the iteration progresses. So, when p is updated to z + βp then
Cp = r + βCp can be updated at the same time without computing the product of C with p.
Then
p
C = pT Cp. Similarly d = d + τ p implies that Cd = Cd + τ Cp.
Algorithm cgtrust combines the solution of the trust region problem from trcg, the
trust region radius adjustment scheme from trtest, and (indirectly) the locally convergent
algorithm newtcg. The result fits nicely into our paradigm algorithm trgen.

Algorithm 3.3.8. cgtrust(x, f, τ )
1. Initialize ∆, M , η, kmax.
2. Do forever
(a) Let xc = x. Compute ∇f (xc ).

(b) Call trcg(d, x, f, M, η, ∆, kmax) to solve the trust region subproblem.
Set xt = x + d.
(c) Call trtest(xc , xt , x, f, ∆),
solving the trust region subproblem with Algorithm trcg.
(d) Update η.
Theorem 3.3.10 combines several results from [247].

Theorem 3.3.10. Let f be twice Lipschitz continuously differentiable. Let M be a given
positive definite matrix and let {ηn } satisfy 0 < ηn < 1 for all n. Let {xn } be the sequence
generated by Algorithm cgtrust and assume that {
∇2 f (xn )
} is bounded. Then
(3.49) lim ∇f (xn ) = 0.

n→∞
Moreover, if x∗ is a local minimizer for which the standard assumptions hold and xn → x∗ ,
then
• if ηn ≤ Kη
∇f (xn )
1 + p.
Finally, there are δ and ∆ such that if
x0 − x∗
≤ δ and ∆0 ≤ ∆ then xn → x∗ .
One can, as we do in the MATLAB code cgtrust, replace the Hessian–vector product
with a difference Hessian. The accuracy of the difference Hessian and the loss of symmetry
present the potential problem that was mentioned in §2.5. Another, very different, approach is
to approximate the exact solution of the trust region subproblem with an iterative method [243].
3.4 Examples
The results we report here used the MATLAB implementations of steepest descent, steep.m,
damped Gauss–Newton, gaussn.m, the dogleg trust region algorithm for Newton’s method,
ntrust.m, and the PCG–dogleg algorithms, cgtrust.m, from the software collection.
Our MATLAB implementation of Algorithm steep guards against extremely poor scaling
and very long steps by setting λ to
(3.50) λ0 = min(1, 100/(1 +

∇f (x)
))
at the beginning of the line search. We invite the reader in Exercise 3.5.3 to attempt the control
example with λ0 = 1.
We not only present plots, which are an efficient way to understand convergence rates, but we
also report counts of function, gradient, and Hessian evaluations and the results of the MATLAB
flops command.

Dogleg Dogleg
5 5
10 10
0
10
Function Value
Gradient Norm
0
10
5
10
5
10
10
10
10 15
10 10
0 10 20 30 0 10 20 30
Steepest Descent Steepest Descent

4 5
10 10
2
10
Function Value
Gradient Norm
0
10
0
10
5
10
2
10
4 10
10 10
0 20 40 60 0 20 40 60
Figure 3.1: Steepest Descent and Newton–Dogleg for Parameter ID Problem
Damped Gauss–Newton Damped Gauss–Newton

2 5
10 10
0 0
10 10
Function Value
Gradient Norm
2 5
10 10
4 10
10 10
6 15
10 10
0 2 4 6 0 2 4 6
Levenberg–Marquardt Levenberg–Marquardt
2 5
10 10
0 0
10 10
Function Value
Gradient Norm
2 5
10 10
4 10
10 10
6 15
10 10
0 5 10 15 0 5 10 15
Figure 3.2: Gauss–Newton and Levenberg–Marquardt for Parameter ID Problem

Dogleg–CG Dogleg–CG
5 7
10 10
6
10
Function Value
Gradient Norm
0
10
5
10
5
10
4
10
10 3
10 10
0 2 4 6 0 2 4 6
Steepest Descent Steepest Descent

5 7
10 10
6
10
Function Value
Gradient Norm
0
10
5
10
5
10
4
10
10 3
10 10
0 20 40 60 0 20 40 60
Figure 3.3: Steepest Descent and Dogleg–CG for Discrete Control Problem

We consider the problem from §2.6.1 except we use the initial data x0 = (5, 5)T . Both the Gauss–
Newton and Newton methods will fail to converge with this initial data without globalization (see
Exercise 3.5.14). Newton’s method has particular trouble with this problem because the Newton
direction is not a descent direction in the early phases of the iteration. The termination criterion
and difference increment for the finite difference Hessian was the same as for the computation
in §2.6.1.
In Figure 3.1 we compare the performance of the Newton dogleg algorithm with the steepest
descent algorithm. Our implementation of the classical dogleg in ntrust uses the standard
values
(3.51) ωdown = .5, ωup = 2, µ0 = µlow = .25, and µhigh = .75.
The plots clearly show the locally superlinear convergence of Newton’s method and the linear
convergence of steepest descent. However, the graphs do not completely show the difference
in computational costs. In terms of gradient evaluations, steepest descent was marginally better
than the Newton dogleg algorithm, requiring 50 gradients as opposed to 55 (which includes those
needed for the 18 difference Hessian evaluations) for the Newton dogleg algorithm. However, the
steepest descent algorithm required 224 function evaluations, while the Newton dogleg needed
only 79. As a result, the Newton dogleg code was much more efficient, needing roughly 5 million
floating point operations instead of the 10 million needed by the steepest descent code.
In Figure 3.2 we plot the performance of the damped Gauss–Newton and Levenberg–
Marquardt algorithms. These exploit the least squares structure of the problem and are locally
superlinearly convergent because this is a zero residual problem. They also show that algo-
rithms that effectively exploit the structure of the least squares problem are much more efficient.
Gauss–Newton required 6 gradient evaluations, 14 function evaluations, and 750 thousand float-
ing point operations, and Levenberg–Marquardt required 12 gradients, 23 functions, and 1.3

million floating point operations.

We consider the discrete control problem from §1.6.1 with N = 400, T = 1, y0 = 0,
L(y, u, t) = (y − 3)2 + .5 ∗ u2 , and φ(y, u, t) = uy + t2 .
We chose the poor initial iterate
u0 (t) = 5 + 300 sin(20πt).
This problem can be solved very efficiently with Algorithm cgtrust. In our implementa-
tion we use the same parameters from (3.51). In Figure 3.3 we compare the dogleg–CG iteration
with steepest descent. We terminated both iterations when
∇f
< 10−8 . For the dogleg–CG
code we used η = .01 throughout the entire iteration and an initial trust region radius of
u0
.
The steepest descent computation required 48 gradient evaluations, 95 function evaluations, and
roughly 1 million floating point operations, and dogleg–CG needed 17 gradient evaluations, 21
function evaluations, and roughly 530 thousand floating point operations. Note that the steepest
descent algorithm performed very well in the terminal phase of the iteration. The reason for this
is that, in this example, the Hessian is near the identity.
3.5 Exercises on Global Convergence

3.5.1. Let F be a nonlinear function from RN → RN . Let
f (x) =
F (x)
2 /2.
What is ∇f ? When is the Newton step for the nonlinear equation F (x) = 0,
d = −F (x)−1 F (x),
a descent direction for f at x?
3.5.2. Prove Lemma 3.2.1.
3.5.3. Implement Algorithm steep without the scaling fixup in (3.50). Apply this crippled
algorithm to the control problem example from §3.4.2. What happens and why?
3.5.4. Show that if f is a convex quadratic then f is bounded from below.
3.5.5. Verify (3.40).
3.5.6. Show that the Levenberg–Marquardt steps computed by (3.20) and (3.21) are the same.
3.5.8. Complete the proof of Theorem 3.3.2.
3.5.10. Look at the trust region algorithm for nonlinear equations from [218] or [84]. What are the
costs of that algorithm that are not present in a line search? When might this trust region
approach have advantages for solving nonlinear equations? Could it be implemented
inexactly?

3.5.11. The double dogleg method [80], [84] puts a new node on the dogleg path in the Newton
direction, thereby trying more aggressively for superlinear convergence. Implement this
method, perhaps by modifying the MATLAB code ntrust.m, and compare the results
with the examples in §3.4. Prove convergence results like Theorems 3.3.7 and 3.3.9 for
this method.
3.5.12. In [51] a trust region algorithm was proposed that permitted inaccurate gradient computa-
tions, with the relative accuracy being tightened as the iteration progresses. Look at [51]
and try to design a similar algorithm based on the line search paradigm. What problems
do you encounter? How do you solve them?
3.5.13. Suppose one modifies Algorithm trtest by not resolving the trust region problem if
the trial point is rejected, but instead performing a line search from xt , and setting ∆ =

x+ − xc
, where x+ is the accepted point from the line search. Discuss the merits of
this modification and any potential problems. See [209] for the development of this idea.
3.5.14. Write programs for optimization that take full Gauss–Newton or Newton steps (you can
cripple the MATLAB codes gaussn.m and ntrust.m for this). Apply these codes to
the parameter identification problem from §3.4.1. What happens?
3.5.15. Write a nonlinear CG code and apply it to the problems in §3.4. Try at least two ways to
manage the line search. How important are the (strong) Wolfe conditions?
3.5.16. Discuss the impact of using a difference Hessian in Algorithm trcg. How will the global
convergence of Algorithm cgtrust be affected? How about the local convergence?
Consider the accuracy in the evaluation of ∇f in your results.
3.5.17. Without looking at [247] describe in general terms how the proof of Theorem 3.3.1 should
be modified to prove Theorem 3.3.10. Then examine the proof in [247] to see if you left
anything out.


Chapter 4
The BFGS Method
Quasi-Newton methods update an approximation of ∇2 f (x∗ ) as the iteration progresses. In

general the transition from current approximations xc and Hc of x∗ and ∇2 f (x∗ ) to new ap-
proximations x+ and H+ is given (using a line search paradigm) by the following steps:
1. Compute a search direction d = −Hc−1 ∇f (xc ).
2. Find x+ = xc + λd using a line search to insure sufficient decrease.
3. Use xc , x+ , and Hc to update Hc and obtain H+ .
The way in which H+ is computed determines the method.
The BFGS (Broyden, Fletcher, Goldfarb, Shanno) [36], [103], [124], [237] method, which
is the focus of this chapter, and the other methods we will mention in §4.3 are also called secant
methods because they satisfy the secant equation
(4.1) H+ s = y.
In (4.1)
s = x+ − xc and y = ∇f (x+ ) − ∇f (xc ).
If N = 1, all secant methods reduce to the classical secant method for the single nonlinear
equation f (x) = 0, i.e.,
f (xc )(xc − x− )
(4.2) x+ = xc − ,
f (xc ) − f (x− )
where x− is the iterate previous to xc .
The standard quasi-Newton update for nonlinear equations is Broyden’s [34] method, a
rank-one update,
(y − Hc s)sT
(4.3) H+ = Hc + .
sT s
Broyden’s method does not preserve the structural properties needed for line search methods in
optimization, namely, symmetry and positive definiteness, and could, in fact, encourage con-
vergence to a local maximum. For that reason quasi-Newton methods in optimization are more
complex than those used for nonlinear equations. The methods of analysis and implementation
are more complex as well.
In this chapter we will concentrate on the BFGS method [36], [103], [124], [237], which is
the rank-two update
yy T (Hc s)(Hc s)T
(4.4) H+ = Hc + T − .
y s sT Hc s
We will briefly discuss other updates and variations that exploit problem structure in §4.3.
71
4.1 Analysis
This section begins with some simple observations on nonsingularity and positivity of the update.
It is very useful for both theory and practice to express (4.4) in terms of the inverse matrices.
The formula we use in this book is Lemma 4.1.1.
−1
Lemma 4.1.1. Let Hc be spd, y T s = 0, and H+ given by (4.4). Then H+ is nonsingular
and
−1 sy T ysT ssT
(4.5) H+ = I− T Hc−1 I − T + T .
y s y s y s
Proof. See exercise 4.5.2.

Lemma 4.1.2. Let Hc be spd, y T s > 0, and H+ given by (4.4). Then H+ is spd.
Proof. Positivity of Hc and y T s = 0 imply that for all z = 0,
(z T y)2 (z T Hc s)2
z T H+ z = T
+ z T Hc z − T .
y s s Hc s
Using the symmetry and positivity of Hc , we have
(z T Hc s)2 ≤ (sT Hc s)(z T Hc z),
with equality only if z = 0 or s = 0, and, therefore, since z, s = 0 and y T s > 0,
(z T y)2
z T H+ z > ≥ 0,
yT s
as asserted.
If y T s ≤ 0 the update is considered a failure.
4.1.1 Local Theory

The local theory [37] requires accurate initial approximations to both x∗ and ∇2 f (x∗ ). The
statement of the convergence result is easy to understand.
Theorem 4.1.3. Let the standard assumptions hold. Then there is δ such that if

x0 − x∗
≤ δ and
H0 − ∇2 f (x∗ )
≤ δ,
then the BFGS iterates are defined and converge q-superlinearly to x∗ .
Technical Details
The proof of Theorem 4.1.3 is technical and we subdivide it into several lemmas. Our proof is
a hybrid of ideas from [37], [135], and [154]. Similar to other treatments of this topic [45] we
begin with the observation (see §2.5.2) that one may assume ∇2 f (x∗ ) = I for the convergence
analysis.
Lemma 4.1.4. Let the standard assumptions hold and let
fˆ(y) = f (Ay),

BFGS METHOD 73
where A = (∇2 f (x∗ ))−1/2 . Let xc and Hc be given and let x̂c = A−1 xc and Ĥc = AHc A.
Then the BFGS updates (x+ , H+ ) for f and (x̂+ , Ĥ+ ) for fˆ are related by
x̂+ = A−1 x+ and Ĥ+ = AH+ A.
In particular, the BFGS sequence for f exists (i.e., Hn is spd for all n) if and only if the BFGS
sequence for fˆ does and the convergence of {xn } is q-superlinear if and only if the convergence
of {x̂n } is.
Proof. The proof is a simple calculation and is left for exercise 4.5.3.
Hence we can, with no loss of generality, assume that ∇2 f (x∗ ) = I, for if this is not so, we
can replace f by fˆ and obtain an equivalent problem for which it is.
Keeping in mind our assumption that ∇2 f (x∗ ) = I, we denote errors in the inverse Hessian
by
E = H −1 − ∇2 f (x∗ )−1 = H −1 − I.
These errors satisfy a simple recursion [37].
Lemma 4.1.5. Let the standard assumptions hold. Let Hc be spd and
x+ = xc − Hc−1 ∇f (xc ).
Then there is δ0 such that if
0 <
xc − x∗
≤ δ0 and
Ec
≤ δ0 ,
then y T s > 0. Moreover, if H+ is the BFGS update of Hc then
(4.6) E+ = (I − wwT )Ec (I − wwT ) + ∆,
where w = s/
s
and for some K∆ > 0
(4.7)
∆
≤ K∆
s
.
Proof. Let δ0 be small enough so that ∇f (xc ) = 0 if xc = x∗ . Theorem 1.2.1 implies that
1
∇f (xc ) = ∇2 f (x∗ + tec )ec dt = ec + ∆1 ec ,
0
where ∆1 is the matrix given by

1
∆1 = (∇2 f (x∗ + tec ) − I) dt.
0
Clearly

∆1
≤ γ
ec
/2,
and
s = −Hc−1 ∇f (xc ) = −(I + Ec )(I + ∆1 )ec .
Therefore,

ec
(1 − δ0 )(1 − γδ0 /2) ≤
s
≤
ec
(1 + δ0 )(1 + γδ0 /2)
and hence
(4.8) 0 <
ec
/2 ≤
s
≤ 2
ec
if, say,
(4.9) δ0 ≤ min(1/4, 1/(2γ)).

We will assume that (4.9) holds for the rest of this section.
The standard assumptions, our assumption that ∇2 f (x∗ ) = I, and the fundamental theorem
of calculus imply that
1
y = ∇f (x+ ) − ∇f (xc ) = ∇2 f (xc + ts)s dt
0
(4.10)
1
=s+ (∇2 f (xc + ts) − I)s dt = s + ∆2 s,
0
where ∆2 is the matrix given by

1
∆2 = (∇2 f (xc + ts) − I) dt.
0
The standard assumptions imply that

∆2
≤ γ(
e+
+
ec
)/2. Hence, (4.8) implies that
(4.11) y T s = sT s + (∆2 s)T s ≥

s
2 (1 − 3γ
ec
/2)
s
2 (1 − 3γδ0 /2) > 0
provided δ0 < 2γ/3. We have that
sy T ssT + s(∆2 s)T ssT

(4.12) = = − ∆3 = wwT − ∆3 ,
yT s sT s + (∆2 s)T s sT s
where (see exercise 4.5.4), for some C > 0,
(4.13)
∆3
≤ C
s
.
Subtracting (∇2 f (x∗ ))−1 = I from (4.5) and using (4.12) gives us
E+ = (I − wwT + ∆3 )Hc−1 (I − wwT + ∆T3 ) + wwT − I
= (I − wwT )(Ec + I)(I − wwT ) + wwT − I + ∆
= (I − wwT )Ec (I − wwT ) + ∆,
where
∆ = ∆3 Hc−1 (I − wwT + ∆T3 ) + (I − wwT )Hc−1 ∆T3 .
Therefore, if (4.9) holds then 1 + δ0 ≤ 3/2 and

∆
≤ (1 + δ0 )
∆3
(2 +
∆3
) ≤
s
3C(2 + C
s
)/2
≤ 3C
s
(2 + 2Cδ0 )/2.
Reduce δ0 if necessary so that 2Cδ0 ≤ 1 and the proof is complete with K∆ = 9C/2.
Lemma 4.1.5 implies that the approximate Hessians do not drift too far from the exact Hessian
if the initial data are good. This property, called bounded deterioration in [37], will directly
imply local q-linear convergence.
Corollary 4.1.6. Let the assumptions of Lemma 4.1.5 hold and let δ0 be as in the statement
of Lemma 4.1.5. Then
(4.14)
E+
≤
Ec
+ K∆
s
≤
Ec
+ K∆ (
ec
+
e+
).

BFGS METHOD 75
Proof. The leftmost inequality follows from Lemma 4.1.5 and the fact that I − wwT is an
orthogonal projection. The final inequality follows from the triangle inequality.
We are now ready to prove local q-linear convergence. This is of interest in its own right and
is a critical step in the superlinear convergence proof. Note that, unlike the statement and proof
of Theorem 2.3.4, we do not express the estimates in terms of
H − ∇2 f (x∗ )
=
H − I
but
in terms of E = H −1 − I. The two approaches are equivalent, since if
E
≤ δ. < 1/2, then
H −1
< 3/2 and the Banach lemma implies that
H
≤ 2. Hence

Hn − I
/2 ≤
Hn
−1
Hn − I
≤
Hn−1 − I
=
Hn−1 (Hn − I)
≤
Hn−1
Hn − I
≤ 3
Hn − I
/2.
Theorem 4.1.7. Let the standard assumptions hold and let σ ∈ (0, 1). Then there is δ. such
that if
(4.15)
x0 − x∗
≤ δ. and
H0−1 − ∇2 f (x∗ )−1
≤ δ. ,
then the BFGS iterates are defined and converge q-linearly to x∗ with q-factor at most σ.
Proof. For δ̂ sufficiently small and
(4.16)
xc − x∗
≤ δ̂ and
Ec
=
Hc−1 − I
≤ δ̂,
the standard assumptions imply that there is K̄ such that
(4.17)
e+
≤ K̄(
Ec
ec
+
ec
2 )/2 ≤ K̄ δ̂
ec
.
Reduce δ̂ if necessary so that K̄ δ̂ ≤ σ to obtain

e+
≤ σ
ec
. The method of proof is to select
δ. so that (4.16) is maintained for the entire iteration if the initial iterates satisfy (4.15).
With this in mind we set
−1
∗ K∆ (1 + σ)
(4.18) δ. = δ /2 1 + < δ ∗ /2
1−σ
where K∆ is from Lemma 4.1.5. Now if
H0 − I
< δ. then

E0
≤ δ. /(1 − δ. ) ≤ 2δ. < δ ∗
which is the estimate we need.
Now by Corollary 4.1.6

E1
≤
E0
+ K∆ (1 + σ)
e0
.
The proof will be complete if we can show that (4.15) and (4.18) imply that
En
< δ ∗ for all
n. We do this inductively. If
En
< δ ∗ and
ej+1
≤ σ
ej
for all j ≤ n, then (4.14) implies
that

En+1
≤
En
+ K∆ (
en
+
en+1
) ≤
En
+ K∆ (1 + σ)
en
≤
En
+ K∆ (1 + σ)σ n
e0
≤
En
+ K∆ (1 + σ)σ n δ.
n
≤
E0
+ δ. K∆ (1 + σ) σn
j=0

K∆ (1 + σ)
= δ. 1 + .
1−σ
We complete the induction and the proof by invoking (4.18) to conclude that
En+1
≤ δ ∗ .

Proof of Theorem 4.1.3

The Dennis–Moré condition [82], [81] is a necessary and sufficient condition for superlinear
convergence of quasi-Newton methods. In terms of the assumptions we make in this section,
the condition is

En sn
(4.19) lim = 0,
n→∞
sn
where {sn } is the sequence of steps and {En } is the sequence of errors in the inverse Hessian.
We will only state and prove the special case of the necessary condition that we need and refer
the reader to [82], [81], [84], or [154] for more general proofs.
Theorem 4.1.8. Let the standard assumptions hold; let {Hn } be a sequence of nonsingular
N × N matrices satisfying
(4.20)
Hn
≤ M
for some M > 0. Let x0 ∈ RN be given and let {xn }∞
n=1 be given by
xn+1 = xn − Hn−1 ∇f (xn )
for some sequence of nonsingular matrices Hn . Then if xn → x∗ q-linearly, xn = x∗ for any

n, and (4.19) holds then xn → x∗ q-superlinearly.
Proof. We begin by invoking (4.10) to obtain
En sn = (Hn−1 − I)sn = (Hn−1 − I)(yn − ∆2 s) = En yn + O(

sn
2 ).
Convergence of xn to x∗ implies that sn → 0 and hence (4.19) can be written as

En yn
(4.21) lim = 0,
n→∞
sn
where yn = ∇f (xn+1 ) − ∇f (xn ).

Now let σ be the q-factor for the sequence {xn }. Clearly
(1 − σ)
en
≤
sn
≤ (1 + σ)
en
.
Hence (4.21) is equivalent to

En yn
(4.22) lim = 0.
n→∞
en
Since Hn−1 ∇f (xn ) = −sn and sn = yn + O(

sn
2 ) we have
En y n = (Hn−1 − I)(∇f (xn+1 ) − ∇f (xn ))
= Hn−1 ∇f (xn+1 ) + sn − yn = Hn−1 ∇f (xn+1 ) + O(

sn
2 )
= Hn−1 en+1 + O(
en
2 +
sn
2 ) = Hn−1 en+1 + O(
en
2 ).
Therefore, (4.22) implies that

En yn

Hn−1 en+1

en+1
= + O(
en
) ≥ M −1 + O(
en
) → 0

en

en

en
as n → ∞, proving q-superlinear convergence.

For the remainder of this section we assume that (4.15) holds and that δ. is small enough so
that the conclusions of Theorem 4.1.7 hold for some σ ∈ (0, 1). An immediate consequence of
this is that ∞
(4.23)
sn
< ∞.
n=0

BFGS METHOD 77
The Frobenius norm of a matrix A is given by

N

(4.24)
A
2F = (A)2ij .
i,j=1
It is easy to show that (see exercise 4.5.5) for any unit vector v ∈ RN ,
(4.25)
A(I − vv T )
2F ≤
A
2F −
Av
2 and
(I − vv T )A
2F ≤
A
2F .
We have, using (4.6), (4.7), and (4.25), that
(4.26)
En+1
2F ≤
En
2F −
En wn
2 + O(
sn
) = (1 − θn2 )
En
2F + O(
sn
),
where wn = sn /
sn
and



En wn
 if En = 0,

En
F
θn =



1 if En = 0.
Using (4.23) we see that for any k ≥ 0,

k k
θn2
En
2F ≤
En
2F −
En+1
2F + O(1)
n=0 n=0
=
E0
2F −
Ek+1
2F + O(1) < ∞.
Hence θn
En
F → 0.
However, 

En wn
if En = 0
θn
En
F =

0 if En = 0

En sn
=
En wn
= .

sn
Hence (4.19) holds. This completes the proof of Theorem 4.1.3.
4.1.2 Global Theory

If one uses the BFGS model Hessian in the context of Algorithm optarm, then Theorem 3.2.4
can be applied if the matrices {Hk } remain bounded and well conditioned. However, even if a
limit point of the iteration is a minimizer x∗ that satisfies the standard assumptions, Theorem 3.2.4
does not guarantee that the iteration will converge to that point. The situation in which x is near
x∗ but H is not near ∇2 f (x∗ ) is, from the point of view of the local theory, no better than that
when x is far from x∗ . In practice, however, convergence (often superlinear) is observed. The
result in this section is a partial explanation of this.
Our description of the global theory, using the Armijo line search paradigm from Chapter 3, is
based on [43]. We also refer the reader to [221], [45], and [269] for older results with a different
line search approach. Results of this type require strong assumptions on f and the initial iterate
x0 , but the reward is global and locally superlinear convergence for a BFGS–Armijo iteration.
Assumption 4.1.1. The set
D = {x | f (x) ≤ f (x0 )}

is convex and f is Lipschitz twice continuously differentiable in D. Moreover, there are λ+ ≥

λ− > 0 such that
σ(∇2 f (x)) ⊂ [λ− , λ+ ]
for all x ∈ D.
Assumption 4.1.1 implies that f has a unique minimizer x∗ in D and that the standard
assumptions hold near x∗ .
Theorem 4.1.9. Let Assumption 4.1.1 hold and let H0 be spd. Then the BFGS–Armijo
iteration converges q-superlinearly to x∗ .
The results for local and global convergence do not completely mesh. An implementation
must allow for the fact that Assumption 4.1.1 may fail to hold, even near the root, and that
y T s ≤ 0 is a possibility when far from the root.
4.2 Implementation
The two implementation issues that we must confront are storage of the data needed to maintain
the updates and a strategy for dealing with the possibility that y T s ≤ 0. We address the
storage question in §4.2.1. For the second issue, when y T s is not sufficiently positive, we restart
the BFGS update with the identity. We present the details of this in §4.2.2. Our globalization
approach, also given in §4.2.2, is the simplest possible, the Armijo rule as described in Chapter 3.
We choose to discuss the Armijo rule in the interest of simplicity of exposition. However,
while the Armijo rule is robust and sufficient for most problems, more complex line search
schemes have been reported to be more efficient [42], and one who seeks to write a general
purpose optimization code should give careful thought to the best way to globalize a quasi-
Newton method. In the case of BFGS, for example, one is always seeking to use a positive definite
quadratic model, even in regions of negative curvature, and in such regions the approximate
Hessian could be reinitialized to the identity more often than necessary.
4.2.1 Storage
For the present we assume that y T s > 0. We will develop a storage-efficient way to compute
the BFGS step using the history of the iteration rather than full matrix storage.
The implementation recommended here is one of many that stores the history of the iteration
and uses that information recursively to compute the action of Hk−1 on a vector. This idea was
suggested in [16], [186], [206], and other implementations may be found in [44] and [201].
All of these implementations store the iteration history in the pairs {sk , yk } and we present a
concrete example in Algorithm bfgsrec. A better, but somewhat less direct, way is based on
the ideas in [91] and [275] and requires that only a single vector be stored for each iteration. We
assume that we can compute the action of H0−1 on a vector efficiently, say, by factoring H0 at the
outset of the iteration or by setting H0 = I. We will use the BFGS formula from Lemma 4.1.1.
One way to maintain the update is to store the history of the iteration in the sequences of
vectors {yk } and {sk } where
sk = xk+1 − xk and yk = ∇f (xk+1 ) − ∇f (xk ).
If one has done this for k = 0, . . . , n − 1, one can compute the new search direction
dn = −Hn−1 ∇f (xn )
by a recursive algorithm which applies (4.5).

BFGS METHOD 79
Algorithm bfgsrec overwrites a given vector d with Hn−1 d. The storage needed is one
vector for d and 2n vectors for the sequences {sk , yk }n−1
k=0 . A method for computing the product
of H0−1 and a vector must also be provided.
Algorithm 4.2.1. bfgsrec(n, {sk }, {yk }, H0−1 , d)
1. If n = 0, d = H0−1 d; return
2. α = sTn−1 d/yn−1
T
s; d = d − αyn−1
3. call bfgsrec(n − 1, {sk }, {yk }, H0−1 , d)

T T
4. d = d + (α − (yn−1 d/yn−1 sn−1 ))sn−1
Algorithm bfgsrec has the great advantage, at least in a language that efficiently supports
recursion, of being very simple. More complex, but nonrecursive versions, have been described
in [16], [201], and [44].
The storage cost of two vectors per iteration can be significant, and when available storage is
exhausted one can simply discard the iteration history and restart with H0 . This approach, which
we implement in the remaining algorithms in this section, takes advantage of the fact that if H0
is spd then −H0−1 ∇f (x) will be a descent direction, and hence useful for a line search. Another
approach, called the limited memory BFGS [44], [207], [176], [201], keeps all but the oldest
(s, y) pair and continues with the update. Neither of these approaches for control of storage,
while essential in practice for large problems, has the superlinear convergence properties that
the full-storage algorithm does.
At a cost of a modest amount of complexity in the formulation, we can reduce the storage
cost to one vector for each iteration. The method for doing this in [275] begins with an expansion
of (4.5) as
−1
H+ = Hc−1 + α0 sc sTc + β0 ((Hc−1 yc )sTc + sc (Hc−1 yc )T ),
where
ycT sc + ycT Hc−1 yc −1
α0 = and β0 = T .
(ycT sc )2 yc sc
Now note that
Hc−1 yc = Hc−1 ∇f (x+ ) − Hc−1 ∇f (xc ) = Hc−1 ∇f (x+ ) + sc /λc
and obtain
−1
(4.27) H+ = Hc−1 + α1 sc sTc + β0 (sc (Hc−1 ∇f (x+ ))T + (Hc−1 ∇f (x+ ))sTc ),
where
α1 = α0 + 2β0 /λc .
Also
−1
d+ = −H+ ∇f (x+ )

sc y T yc sT sc sTc ∇f (x+ )
(4.28) =− I− T c Hc−1 I− Tc ∇f (x+ ) −
yc sc yc sc ycT sc
= Ac sc + Bc Hc−1 ∇f (x+ ),
where
ycT yc sTc sT ∇f (x+ )
(4.29) Ac = T
H −1
c I − T
∇f (x+ ) + c T
yc sc yc sc λc yc sc

and
sTc ∇f (x+ )
(4.30) Bc = −1 + .
ycT sc
At this point we can compute d+ , and therefore λ+ and s+ using only Hc−1 ∇f (xc ). We do not
need H+ at all. We can now form H+ with the new data for the next iterate and will show that
we do not need to store the vectors {yk }.
Since (verify this!) Bc = 0, we have
s+ Ac sc
Hc−1 ∇f (x+ ) = − + .
Bc λ+ Bc
Combining this with (4.27) gives

−1
(4.31) H+ = Hc−1 + αc sc sTc + βc (sc sT+ + s+ sTc ),
where
β0
(4.32) αc = α1 + 2β0 Ac /Bc and βc = − .
Bc λ+
This leads to the expansion
n

−1
(4.33) Hn+1 = H0−1 + αk sk sTk + βk (sk sTk+1 + sk+1 sTk ).
k=0
Upon reflection the reader will see that this is a complete algorithm. We can use (4.28) and Hn
to compute dn+1 . Then we can compute λn+1 and sn+1 and use them and (4.32) to compute
−1
αn and βn . This new data can be used to form Hn+1 with (4.33), which we can use to compute
dn+2 and continue the iteration.
In this way only the steps {sk } and the expansion coefficients {αk } and {βk } need be stored.
Algorithm bfgsopt is an implementation of these ideas.
Algorithm 4.2.2. bfgsopt(x, f, )
1. g = −∇f (x), n = 0.
2. While
g
>
(a) If n = 0, dn = −H0−1 g
otherwise compute A, B, and dn using (4.28), (4.29), and (4.30).
(b) Compute λn , sn , and x = xn+1 with the Armijo rule.
(c) If n > 0 compute αn−1 and βn−1 using (4.32).
(d) g = −∇f (x), n = n + 1.
4.2.2 A BFGS–Armijo Algorithm

In this section we present a simple implementation that shows how the theoretical results can
BF GS
be applied in algorithm design. Let H+ be the BFGS update from Hc and define the two
modified BFGS (MBFGS) updates by
BF GS
H+ if y T s > 0,
(4.34) H+ =
I if y T s ≤ 0,

BFGS METHOD 81
and BF GS
H+ if y T s > 0,
(4.35) H+ =
Hc if y T s ≤ 0.
In the MBFGS1 method, (4.34), the model Hessian is reinitialized to I if y T s ≤ 0. In the
early phase of this iteration, where ∇2 f may have negative eigenvalues, y T s ≤ 0 is certainly
possible and the search direction could be the steepest descent direction for several iterations.
An MBFGS2 step (4.35) keeps the history of the iteration even if y T s ≤ 0. One view
is that this approach keeps as much information as possible. Another is that once y T s ≤ 0,
the iteration history is suspect and should be thrown away. Both forms are used in practice.
Our MATLAB code bfgswopt uses MFBGS1 and maintains an approximation to H −1 using
Algorithm bfgsopt. We also guard against poor scaling by using (3.50).
4.3 Other Quasi-Newton Methods

The DFP (Davidon, Fletcher, Powell) update [71], [72], [105]
(y − Hc s)y T + y(y − Hc s)T [(y − Hc s)T y]yy T

(4.36) H + = Hc + −
yT s (y T s)2
has similar local convergence properties to BFGS but does not perform as well in practice [224],
[225].
Two updates that preserve symmetry, but not definiteness, are the PSB (Powell symmetric
Broyden) update [219],
(y − Hc s)sT + s(y − Hc s)T [sT (y − Hc s)]ssT

(4.37) H + = Hc + − ,
sT s (sT s)2
and the symmetric rank-one (SR1) [35] update,
(y − Hc s)(y − Hc s)T
(4.38) H+ = Hc + .
(y − Hc s)T s
By preserving the symmetry of the approximate Hessians, but not the positive definiteness,
these updates present a problem for a line search globalization but an opportunity for a trust
region approach. The SR1 update has been reported to outperform BFGS algorithms in certain
cases [165], [41], [64], [65], [163], [258], [118], [250], [119], [268], [164], in which either the
approximate Hessians can be expected to be positive definite or a trust region framework is used
[41], [64], [65].
One may update the inverse of the SR1 approximate Hessian using the Sherman–Morrison
formula, (4.39), a simple relation between the inverse of a nonsingular matrix and that of a
rank-one update of that matrix [93], [239], [240], [14].
Proposition 4.3.1. Let H be a nonsingular N ×N matrix and let u, v ∈ RN . Then H +uv T
is invertible if and only if 1 + v T H −1 u = 0. In this case

(H −1 u)v T
(4.39) (H + uv T )−1 = I − H −1 .
1 + v T H −1 u
The proof is simply a direct verification of (4.39).

The SR1 algorithm terminates in finitely many iterations for convex quadratic optimization
problems [101]. Since the denominator (y − Hc s)T s could vanish, the update could completely
fail and implementations must examine the denominator and take appropriate action if it is too

small. This update does not enforce or require positivity of the approximate Hessian and has
been used effectively to exploit negative curvature in a trust region context [165], [41].
For overdetermined nonlinear least squares problems one can try to approximate the second-
order term in ∇2 f while computing RT R exactly. Suppose
∇2 f (x) ≈ H = C(x) + A,
where the idea is that C, the computed part, is significantly easier to compute than A, the
approximated part. This is certainly the case for nonlinear least squares, where C = RT R . A
quasi-Newton method that intends to exploit this structure will update A only; hence
H+ = C(x+ ) + A+ .
Superlinear convergence proofs require, in one way or another, that H+ s = y. Therefore, in

terms of A, one might require the update to satisfy
(4.40) A+ s = y # = y − C(x+ )s.
The definition of y # given in (4.40) is called the default choice in [87]. This is not the only
choice for y # , and one can prove superlinear convergence for this and many other choices [87],
[84]. This idea, using several different updates, has been used in other contexts, such as optimal
control [159], [164].
An algorithm of this type, using SR1 to update A and a different choice for y # , was suggested
in [20] and [21]. The nonlinear least squares update from [77], [78], and [84] uses a DFP update
and yet another y # to compute A+ ,
T T
(y # − Ac s)y # + y # (y # − Ac s)T [(y # − Ac s)T y # ]y # y #
(4.41) A+ = Ac + T
− T
.
y# s (y # s)2
The application of this idea to large-residual least squares problems is not trivial, and scaling
issues must be considered in a successful implementation.
Our proof of superlinear convergence can be applied to updates like (4.41). We state a special
case of a result from [87] for the BFGS formulation
T
y# y# (Ac s)(Ac s)T
(4.42) A+ = Ac + − .
y #T s sT Ac s
Theorem 4.3.2. Let the standard assumptions hold and assume that
A∗ = ∇2 f (x∗ ) − C(x∗ )
is spd. Then there is δ such that if

x0 − x∗
≤ δ and
A0 − A∗
≤ δ,
then the quasi-Newton iterates defined by (4.42) exist and converge q-superlinearly to x∗ .
This result can be readily proved using the methods in this chapter (see [159]).
Quasi-Newton methods can also be designed to take into account special structure, such as
the sparsity pattern of the Hessian. One can update only those elements that are nonzero in the
initial approximation to the Hessian, requiring that the secant equation Hs = y holds. Such
updates have been proposed and analyzed in varying levels of generality in [83], [87], [185],
[238], [256], and [255].

BFGS METHOD 83
Another approach is to use the dependency of f on subsets of the variables, a structure that is
often present in discretizations of infinite-dimensional problems where coefficients of operators
can be updated rather than entire matrix representations of those operators. We refer the reader
to [133], [131], and [132] for an algebraic viewpoint based on finite-dimensional analysis and
to [159], [157], [164], [160], and [163] for an operator theoretic description of these methods.
When applied to discretizations of infinite-dimensional optimization problems, quasi-Newton
methods perform best when they also work well on the infinite-dimensional problem itself. Work
on BFGS in Hilbert space can be found, for example, in [135], [158], and [159].
Quasi-Newton methods have been designed for underdetermined problems [184], and Broy-
den’s method itself has been applied to linear least squares problems [111], [148].
4.4 Examples
The computations in this section were done with the MATLAB code bfgswopt. For the small
parameter ID problem, where evaluation of f is far more expensive than the cost of maintaining
or factoring the (very small!) approximate Hessian, one could also use a brute force approach
in which H is updated and factored anew with each iteration.
4.4.1 Parameter ID Problem

We solve the parameter ID problem with the same data as in §3.4.1 using H0 = I as the initial
Hessian. We compare the BFGS solution with the Gauss–Newton iteration from §3.4.1. From
Figure 4.1 one can see the local superlinear convergence and the good performance of the line
search. However, as one should expect, the Gauss–Newton iteration, being designed for small
residual least squares problems, was more efficient. The Gauss–Newton iteration required 14
function evaluations, 6 gradients, and roughly 1.3 million floating point operations, while the
BFGS–Armijo iteration needed 29 function evaluations, 15 gradients, and 3.8 million floating
point operations.

We return to the example from §3.4.2. For our first example we use the initial iterate
u0 (t) = 10.
BFGS also requires an initial approximation to the Hessian and we consider two such approxi-
mations:
(4.43) Hp = .25I and Hg = I.
The Hessian for the continuous problem is a compact perturbation of the identity and the
theory from [158] and [135] indicates that Hg is a much better approximate Hessian than Hp .
The results in Figure 4.2 support that idea. For the better Hessian, one can see the concavity
of superlinear convergence in the plot of the gradient norm. The computation for the better
Hessian required 12 iterations and roughly 572 thousand floating point operations, while the one
with the poor Hessian took 16 iterations and roughly 880 thousand floating point operations.
Stepsize reductions were not required for the good Hessian and were needed four times during
the iteration for the poor Hessian. However, the guard against poor scaling (3.50) was needed
in both cases.
When we used the same poor initial iterate that we used in §3.4.2
u0 (t) = 5 + 300 sin(20πt)

Gauss Newton Gauss Newton

2 5
10 10
0 0
10 10
Function Value
Gradient Norm
2 5
10 10
4 10
10 10
6 15
10 10
0 2 4 6 0 2 4 6
BFGS BFGS
5 5
10 10
0
10
Function Value
Gradient Norm
0 5
10 10
10
10
5 15
10 10
0 5 10 15 0 5 10 15
Figure 4.1: BFGS–Armijo and Gauss–Newton for the Parameter ID Problem
Better Hessian Better Hessian

5 5
10 10
Function Value
Gradient Norm
0
10
4
10
5
10
10 3
10 10
0 5 10 15 0 5 10 15
Poor Hessian Poor Hessian

5 5
10 10
Function Value
Gradient Norm
0
10
4
10
5
10
10 3
10 10
0 5 10 15 20 0 5 10 15 20
Figure 4.2: BFGS–Armijo for Discrete Control Problem

BFGS METHOD 85
and allocated 50 vectors to Algorithm bfgsopt, there was no longer a benefit to using the
good Hessian. In fact, as is clear from Figure 4.3 the poor Hessian produced a more rapidly
convergent iteration.
Better Hessian Better Hessian

10 8
10 10
5
10
Function Value
Gradient Norm
6
10
0
10
4
10
5
10
10 2
10 10
0 50 100 150 0 50 100 150
Poor Hessian Poor Hessian

10 8
10 10
5
10
Function Value
Gradient Norm
6
10
0
10
4
10
5
10
10 2
10 10
0 20 40 60 80 0 20 40 60 80
Figure 4.3: BFGS–Armijo for Discrete Control Problem: Poor Initial Iterate
4.5 Exercises on BFGS

4.5.1. Use the secant method (4.2) with initial data x−1 = 1 and x0 = .9 to minimize f (x) = x4 .
Explain the convergence of the iteration.
4.5.2. Prove Lemma 4.1.1. It might help to use the secant equation.
4.5.4. Verify (4.13) and compute the constant C.
4.5.5. Prove (4.25).
4.5.6. As an exercise in character building, implement Algorithm bfgsrec nonrecursively.
4.5.7. Show how the Sherman–Morrison formula can be used to implement the SR1 update in
such a way that only one vector need be stored for each iterate.
4.5.8. State and prove a local convergence theorem for DFP and/or PSB.
4.5.9. Implement the DFP and PSB update and compare their performance with BFGS on the
examples from §4.4.

4.5.10. Show that, for positive definite quadratic problems, the BFGS method with an exact line
search (i.e., one that finds the minimum of f in the search direction) is the same as CG
[201], [200].

Chapter 5
Simple Bound Constraints
5.1 Problem Statement

The goal of this chapter is to show how the techniques of Chapters 2, 3, and 4 can be used to
solve a simple constrained optimization problem. The algorithm we suggest at the end in §5.5.3
is a useful extension of the BFGS–Armijo algorithm from Chapter 4. We will continue this line
of development when we solve noisy problems in Chapter 7.
Let {Li }N N
i=1 and {Ui }i=1 be sequences of real numbers such that
(5.1) −∞ < Li < Ui < +∞.
The bound constrained optimization problem is to find a local minimizer x∗ of a function f of

N variables subject to the constraint that
(5.2) x∗ ∈ Ω = {x ∈ RN | Li ≤ (x)i ≤ Ui }.
By this we mean that x∗ satisfies
(5.3) f (x∗ ) ≤ f (x) for all x ∈ Ω near x∗ .
It is standard to express this problem as
(5.4) min f (x)

x∈Ω
or as minΩ f . The set Ω is called the feasible set and a point in Ω is called a feasible point.
Because the set Ω is compact there is always a solution to our minimization problem [229].
The inequalities Li ≤ (x)i ≤ Ui are called inequality constraints or simply constraints.
We will say that the ith constraint is active at x ∈ Ω if either (x)i = Li or (x)i = Ui . If the
ith constraint is not active we will say that it is inactive. The set of indices i such that the ith
constraint is active (inactive) will be called the set of active (inactive) indices at x.
We will write A(x) and I(x) for the active and inactive sets at x.
5.2 Necessary Conditions for Optimality

For a continuously differentiable function of one variable, the necessary conditions for uncon-
strained optimality at x∗ are simply f (x∗ ) = 0 and, if f is twice continuously differentiable,
f (x∗ ) ≥ 0. A bound constrained problem in one variable restricts the domain of f to an
interval [a, b], and the necessary conditions must be changed to admit the possibility that the
87
minimizer is one of the endpoints. If x∗ = a is a local minimizer, then it need not be the case
that f (a) = 0; however, because a is a local minimizer, f (x) ≥ f (a) for all a ≤ x sufficiently
near a. Hence f (a) ≥ 0. Nothing, however, can be said about f . Similarly, if x∗ = b is a
local minimizer, f (b) ≤ 0. If f is differentiable on [a, b] (i.e., on an open set containing [a, b]),
then the necessary conditions for all three possibilities, x∗ = a, x∗ = b, and a < x∗ < b can be
neatly expressed by the following theorem.
Theorem 5.2.1. Let f be a continuously differentiable function of one variable on the
interval [a, b]. Let x∗ be a local minimum of f on [a, b]. Then
(5.5) f (x∗ )(x − x∗ ) ≥ 0 for all x ∈ [a, b]
and, if f is twice continuously differentiable on [a, b],
(5.6) f (x∗ )(x∗ − a)(b − x∗ ) ≥ 0.
The analogue (5.5) is expressed by the idea of stationarity.

Definition 5.2.1. A point x∗ ∈ Ω is stationary for problem (5.4) if
(5.7) ∇f (x∗ )T (x − x∗ ) ≥ 0 for all x ∈ Ω.
As in the unconstrained case, stationary points are said to satisfy the first-order necessary
conditions.
The fact that optimality implies stationarity is proved with Taylor’s theorem just as it was in
the unconstrained case.
Theorem 5.2.2. Let f be continuously differentiable on Ω and let x∗ be a solution of problem
(5.4). Then x∗ is a stationary point for problem (5.4).
Proof. Let x∗ be a solution of problem (5.4) and let y ∈ Ω. As Ω is convex, the line segment
joining x∗ and y is entirely in Ω. Hence, the function
φ(t) = f (x∗ + t(y − x∗ ))
is defined for t ∈ [0, 1] and has a local minimum at t = 0. Therefore, by Theorem 5.2.1
0 ≤ φ (0) = ∇f (x∗ )T (y − x∗ )
as asserted.
The case of a function of a single variable is less useful in explaining the role of the second
derivative. However, we can get a complete picture by looking at functions of two variables.
To illustrate the ideas we let N = 2 and let f be a twice Lipschitz continuously differentiable
function on Ω = [0, 1] × [0, 1]. If x∗ is a solution of (5.4) and no constraints are active, then
∇2 f (x∗ ) is positive semidefinite by the same arguments used in the unconstrained case. If one
or more constraints are active, however, then, just as in the one variable case, one cannot draw
conclusions about the positivity of ∇2 f (x∗ ). Suppose the minimizer is at x∗ = (ξ, 0) with
0 < ξ < 1. While nothing can be said about ∂ 2 f (x∗ )/∂x22 , the function φ(t) = f (t, 0), defined
on [0, 1], must satisfy
φ (ξ) = ∂ 2 f (x∗ )/∂x21 ≥ 0.
Hence, second partials in directions corresponding to the inactive constraints must be nonnega-
tive, while nothing can be said about directions corresponding to active constraints.

BOUND CONSTRAINTS 89
We define the reduced Hessian to help make this idea precise.

Definition 5.2.2. Let f be twice differentiable at x ∈ Ω. The reduced Hessian ∇2R f (x) is
the matrix whose entries are

 δij if i ∈ A(x) or j ∈ A(x),
(5.8) (∇2R f (x))ij =

(∇2 f (x))ij otherwise.
We can now present the second-order necessary conditions.

Theorem 5.2.3. Let f be twice Lipschitz continuously differentiable and let x∗ be the solution
of problem (5.4). Then the reduced Hessian ∇2R f (x∗ ) is positive semidefinite.
Proof. Assume that there are M inactive indices and N − M active indices. We partition
x ∈ Ω, reordering the variables if needed, into x = (ξ, ζ) with ξ corresponding to the inactive
indices and ζ to the active. The map
φ(ξ) = f (ξ, ζ ∗ )
has an unconstrained local minimizer at ξ ∗ ∈ RM and hence ∇2 φ is positive semidefinite. Since

the reduced Hessian can be written as
2
∇ φ(x∗ ) 0
∇2R f (x∗ ) =
0 I
if the variables are partitioned in this way, the proof is complete.

We let P denote the projection onto Ω, that is, the map that takes x into the nearest point (in
the l2 -norm) in Ω to x. We have that

 Li if (x)i ≤ Li ,
(5.9) P(x)i = (x)i if Li < (x)i < Ui ,

Ui if (x)i ≥ Ui .
Theorem 5.2.4 states our final necessary condition; we defer the proof to §5.4.4.
Theorem 5.2.4. Let f be continuously differentiable. A point x∗ ∈ Ω is stationary for
problem (5.4) if and only if
(5.10) x∗ = P(x∗ − λ∇f (x∗ ))
for all λ ≥ 0.
5.3 Sufficient Conditions

With the definition of the reduced Hessian in hand, the sufficient conditions are easy to formulate.
We begin by strengthening the notion of stationarity. If x∗ is stationary, i ∈ I(x∗ ), and ei is a
unit vector in the ith coordinate direction, then x∗ ± tei ∈ Ω for all t sufficiently small. Since
df (x∗ ± tei )
= ±∇f (x∗ )T ei ≥ 0,
dt
therefore
(∇f (x∗ ))i = 0 for all i ∈ I(x∗ ).
We will use the concept of nondegenerate stationary point or strict complementarity in our
formulation of the sufficient conditions.

Definition 5.3.1. A point x∗ ∈ Ω is a nondegenerate stationary point for problem (5.4) if

∗
x is a stationary point and
(5.11) (∇f (x∗ ))i = 0 for all i ∈ A(x∗ ).
If x∗ is also a solution of problem (5.4) we say that x∗ is a nondegenerate local minimizer.

Our nondegeneracy condition is also referred to as strict complementarity.
If S is any set of indices define

 (x)i , i ∈ S,
(PS x)i =

0, i ∈ S.
Nondegeneracy is important not only in the formulation of sufficient conditions but also in
the design of termination criteria. The first step in the use of nondegeneracy is Lemma 5.3.1.
Lemma 5.3.1. Let x∗ be a nondegenerate stationary point. Assume that A = A(x∗ ) is not
empty. Then there is σ such that
∇f (x∗ )T (x − x∗ ) = ∇f (x∗ )T PA (x − x∗ ) ≥ σ
PA (x − x∗ )
for all x ∈ Ω.
Proof. If i ∈ A then nondegeneracy and stationarity imply that there is σ > 0 such that
either
(x∗i ) = Li and (∇f (x∗ ))i ≥ σ or (x∗i ) = Ui and (∇f (x∗ ))i ≤ −σ.
If x ∈ Ω then for all i ∈ A,
(∇f (x∗ ))i (x − x∗ )i ≥ σ|(x − x∗ )i |.
Therefore, since
x
1 ≥
x
2 ,
∇f (x∗ )T PA (x − x∗ ) ≥ σ
PA (x − x∗ )
,
as asserted.
For a nondegenerate stationary point the sufficiency conditions are very similar to the un-
constrained case.
Theorem 5.3.2. Let x∗ ∈ Ω be a nondegenerate stationary point for problem (5.4). Let
f be twice differentiable in a neighborhood of x∗ and assume that the reduced Hessian at x∗
is positive definite. Then x∗ is a solution of problem (5.4) (and hence a nondegenerate local
minimizer).
Proof. Let x ∈ Ω and define φ(t) = f (x∗ + t(x − x∗ )). We complete the proof by showing
that either (i) φ (0) > 0 or (ii) φ (0) = 0, φ (0) > 0. Let e = x − x∗ and note that
φ (0) = ∇f (x∗ )T e = ∇f (x∗ )T (PA e + PI e).
Stationarity implies that ∇f (x∗ )T PI e = 0. If PA e = 0 then nondegeneracy implies that
∇f (x∗ )T PA e > 0
and hence (i) holds. If PA e = 0 then
φ (0) = (x − x∗ )T PI ∇2 f (x∗ )PI (x − x∗ ) = (x − x∗ )T ∇2R f (x∗ )(x − x∗ ) > 0,
proving (ii).

5.4 The Gradient Projection Algorithm

The gradient projection algorithm is the natural extension of the steepest descent algorithm to
bound constrained problems. It shares all the advantages and disadvantages of that algorithm.
Our approach follows that of [18]. Given a current iterate xc the new iterate is
x+ = P(xc − λ∇f (xc )),
where λ is a steplength parameter given by the Armijo rule or some other line search scheme.
In this section we will restrict our attention to the simplest form of the Armijo rule. In order to
implement any line search scheme, we must specify what we mean by sufficient decrease. For
λ > 0 define
(5.12) x(λ) = P(x − λ∇f (x)).
For bound constrained problems we will express the sufficient decrease condition for line searches
(compare with (3.4)) as
−α
(5.13) f (x(λ)) − f (x) ≤
x − x(λ)
2 .
λ
As with (3.4), α is a parameter and is typically set to 10−4 [84].
The general algorithmic description follows in Algorithm 5.4.1.
Algorithm 5.4.1. gradproj(x, f, nmax)
1. For n = 1, . . . , nmax

(b) Find the least integer m such that (5.13) holds for λ = β m .
(c) x = x(λ).
2. If n = nmax and the termination test is failed, signal failure.
The next step is to elaborate on the termination criterion.
5.4.1 Termination of the Iteration

The termination criterion for unconstrained optimization that we have used previously must be
modified if we are to properly take the constraints into account. ∇f need not be zero at the
solution, but a natural substitute is to terminate the iteration if the difference between x and x(1)
is small. As in the case of unconstrained optimization or nonlinear equations, we must invoke
the sufficient conditions to show that such a termination criterion will accurately measure the
error.
As usual, we let e = x − x∗ .
We begin with a lemma that connects the active and inactive sets at a nondegenerate local
minimizer with nearby points.
Lemma 5.4.1. Let f be twice continuously differentiable on Ω and let x∗ be a nondegenerate
stationary point for problem (5.4). Let λ ∈ (0, 1]. Then for x sufficiently near x∗ ,
1. A(x) ⊂ A(x∗ ) and (x)i = (x∗ )i for all i ∈ A(x).
2. A(x(λ)) = A(x∗ ) and (x(λ))i = (x∗ )i for all i ∈ A(x∗ ).

Proof. Let
A∗ = A(x∗ ), I ∗ = I(x∗ ), A = A(x), and I = I(x).
Let
δ1 = min∗ {(Ui − (x∗ )i ), ((x∗ )i − Li ), (Ui − Li )/2}.
i∈I
If i ∈ I ∗ and
e
< δ1 then Li < (x)i < Ui . Hence,
I∗ ⊂ I
proving the first assertion that A ⊂ A∗ . Moreover, since

e
< δ1 ≤ min {(Ui − Li )/2} ,
then (x)i = (x∗ )i for all i ∈ A.

Now let Aλ and I λ be the active and inactive sets for x(λ) = P(x − λ∇f (x)). Let i ∈ A∗ .
By Lemma 5.3.1 and continuity of ∇f there is δ2 such that if
e
< δ2 then
(∇f (x∗ + e))i (x − x∗ )i > σ|x − x∗ |i /2.
Therefore, if

e
< δ3 < min(σ/2, δ2 ),
then i ∈ Aλ and (x(λ))i = (x∗ )i . Hence A∗ ⊂ Aλ .
It remains to prove that Aλ ⊂ A∗ . By definition of P we have

P(x) − P(y)
≤
x − y
for all x, y ∈ RN . Continuity of ∇2 f implies that ∇f is Lipschitz continuous. We let L denote

the Lipschitz constant of ∇f in Ω. By stationarity and Theorem 5.2.4,
x∗ = x∗ (λ) = P(x∗ − λ∇f (x∗ )),
and, therefore,

x∗ − x(λ)
=
P(x∗ − λ∇f (x∗ )) − P(x − λ∇f (x))
(5.14)
≤
e
+ λ
∇f (x∗ ) − ∇f (x)
≤ (1 + Lλ)
e
.
If there is i ∈ Aλ ∩ I ∗ then we must have
(5.15)
x∗ − x(λ)
≥ δ1 = min∗ {(Ui − x∗ ), (x∗ − Li )}.
i∈I
However, if

e
< δ4 = min(δ3 , δ1 /(1 + L))
then (5.14) implies that (5.15) cannot hold. This completes the proof.
Theorem 5.4.2. Let f be twice continuously differentiable on Ω and let x∗ be a nondegenerate
stationary point for problem (5.4). Assume that sufficient conditions hold at x∗ . Then there are
δ and M such that if
e
< δ and A(x) = A(x∗ ) then
(5.16)
e
/M ≤
x − x(1)
≤ M
e
.

Proof. Again we let L denote the Lipschitz constant of ∇f in Ω and
A∗ = A(x∗ ), I ∗ = I(x∗ ), A = A(x), and I = I(x).
Using stationarity we obtain

x − x(1)
=
e − (x(1) − x∗ (1))
≤
e
+
P(x − ∇f (x)) − P(x∗ − ∇f (x∗ ))
≤ 2
e
+
∇f (x) − ∇f (x∗ )
≤ (2 + L)
e
.
Hence, the right inequality in (5.16) holds.
To verify the left inequality in (5.16) we apply Lemma 5.4.1. Let δ1 be such that
e
< δ1
implies that the conclusions of Lemma 5.4.1 hold for λ = 1. The lemma implies that

 (∇f (x))i , i ∈ I ∗ ,
(5.17) (x − x(1))i =

(e)i = 0, i ∈ A∗ .
The remaining case is if i ∈ I = I ∗ . The sufficiency conditions imply that there is µ > 0
such that
uT PI ∗ ∇2 f (x∗ )PI ∗ u ≥ µ
PI ∗ u
2
for all u ∈ RN . Hence, there is δ2 so that if
e
< δ2 then
uT PI ∗ ∇2 f (x)PI ∗ u ≥ µ
PI ∗ u
2 /2
for all u ∈ RN .
Therefore, since e = PI ∗ e,
1

PI ∗ (x − x(1))
2 = eT PI ∗ ∇2 f (x∗ + te)e dt
0
1
= eT PI ∗ ∇2 f (x∗ + te)PI ∗ e dt
0
≥ µ
PI∗ e
2 /2.

Therefore,
x − x(1)
≥ min(1, µ/2)
e
and setting M = max{2 + L, 1, 2/µ} completes
the proof.
Following the unconstrained case, we formulate a termination criterion based on relative and
absolute reductions in the measure of stationarity
x − x(1)
. Given r0 =
x0 − x0 (1)
and
relative and absolute tolerances τr and τa the termination criterion for Algorithm gradproj is
(5.18)
x − x(1)
≤ τa + τr r0 .
5.4.2 Convergence Analysis

The convergence analysis is more complicated than that for the steepest descent algorithm be-
cause of the care that must be taken with the constraints. Our analysis begins with several
preliminary lemmas.
Lemma 5.4.3. For all x, y ∈ Ω,
(5.19) (y − x(λ))T (x(λ) − x + λ∇f (x)) ≥ 0.

Proof. By definition of P

x(λ) − x + λ∇f (x)
≤
y − x + λ∇f (x)
for all y ∈ Ω. Hence t = 0 is a local minimum for
φ(t) =
(1 − t)x(λ) + ty − x + λ∇f (x)
2 /2
and, therefore,
0 ≤ φ (0) = (y − x(λ))T (x(λ) − x + λ∇f (x))
as asserted.
We will most often use the equivalent form of (5.19)
(5.20) (x − x(λ))T (y − x(λ)) ≤ λ∇f (x)T (y − x(λ)).
Setting y = x in (5.20), we state Corollary 5.4.4.

Corollary 5.4.4. For all x ∈ Ω and λ ≥ 0,
(5.21)
x − x(λ)
2 ≤ λ∇f (x)T (x − x(λ)).
An important result in any line search analysis is that the steplengths remain bounded away
from 0.
Theorem 5.4.5. Assume that ∇f is Lipschitz continuous with Lipschitz constant L. Let
x ∈ Ω. Then the sufficient decrease condition (5.13) holds for all λ such that
2(1 − α)
(5.22) 0<λ≤ .
L
Proof. We begin with the fundamental theorem of calculus. Setting y = x − x(λ) we have
1
f (x − y) − f (x) = f (x(λ)) − f (x) = − ∇f (x − ty)T y dt.
0
Hence,
f (x(λ)) = f (x) + ∇f (x)T (x(λ) − x)
(5.23) 1
− (∇f (x − ty) − ∇f (x))T y dt.
0
Rearranging terms in (5.23) gives
(5.24) λ(f (x) − f (x(λ))) = λ∇f (x)T (x − x(λ)) + λE,
where 1
E= (∇f (x − ty) − ∇f (x))T y dt
0
and hence

E
≤ L
x − x(λ)
2 /2.

So
(5.25) λ(f (x) − f (x(λ))) ≥ λ∇f (x)T (x − x(λ)) − λL
x − x(λ)
2 /2.
Therefore, using Corollary 5.4.4 we obtain
λ(f (x) − f (x(λ))) ≥ (1 − λL/2)

x − x(λ)
2
which completes the proof.

The consequence for the Armijo rule is that the line search will terminate when
2(1 − α)
βm ≤ ≤ β m−1
L
if not before. Hence, a lower bound for the steplengths is
2β(1 − α)
(5.26) λ̄ = .
L
Theorem 5.4.6. Assume that ∇f is Lipschitz continuous with Lipschitz constant L. Let
{xn } be the sequence generated by the gradient projection method. Then every limit point of
the sequence is a stationary point.
Proof. Since the sequence {f (xn )} is decreasing and f is bounded from below on Ω, f (xn )
has a limit f ∗ . The sufficient decrease condition, as in the proof of Theorem 3.2.4, and (5.26)
imply that

xn − xn+1
2 ≤ λ(f (xn ) − f (xn+1 ))/α ≤ (f (xn ) − f (xn+1 ))/α → 0
as n → ∞.
Now let y ∈ Ω and n ≥ 0. By (5.20) we have
∇f (xn )T (xn − y) = ∇f (xn )T (xn+1 − y) + ∇f (xn )T (xn − xn+1 )
≤ λ−1 T T
n (xn − xn+1 ) (xn+1 − y) + ∇f (xn ) (xn − xn+1 ).
Therefore, by (5.26),
∇f (xn )T (xn − y) ≤
xn − xn+1
(λ−1
n
xn+1 − y
+
∇f (xn )
),
(5.27)
∇f (xn )T (xn − y) ≤
xn − xn+1
(λ̄−1
xn+1 − y
+
∇f (xn )
).
If xnl → x∗ is a convergence subsequence of {xn }, then we may take limits in (5.27) as

l → ∞ and complete the proof.
5.4.3 Identification of the Active Set

The gradient projection iteration has the remarkable property that if it converges to a nondegen-
erate local minimizer, then the active set An of xn is the same as A∗ after only finitely many
iterations.
Theorem 5.4.7. Assume that f is Lipschitz continuously differentiable and that the gradient
projection iterates {xn } converge to a nondegenerate local minimizer x∗ . Then there is n0 such
that A(xn ) = A(x∗ ) for all n ≥ n0 .
Proof. Let λ̄ be the lower bound for the steplength. Let δ be such that the conclusions of
Lemma 5.4.1 hold for λ = λ̄ (and hence for all λ ≥ λ̄). Let n0 be such that
en
< δ for all
n ≥ n0 − 1 and the proof is complete.

5.4.4 A Proof of Theorem 5.2.4

We close this section with a proof of Theorem 5.2.4. We define a nonsmooth function
(5.28) F (x) = x − P(x − ∇f (x)).
Using (5.12),
F (x) = x − x(1).
We now prove Theorem 5.2.4.
Proof. Corollary 5.4.4 states that

x∗ − x∗ (λ)
2 ≤ λ∇f (x∗ )T (x∗ − x∗ (λ)).
If we set x = x∗ (λ) in the definition of stationarity (5.7) we have
∇f (x∗ )T (x∗ − x∗ (λ)) ≤ 0
and hence x∗ = x∗ (λ).

Conversely assume that x∗ = x∗ (λ) for all λ > 0. This implies that x∗ is left invariant by
the gradient projection iteration and is therefore a stationary point.
By setting λ = 1 we obtain a simple consequence of Theorem 5.2.4.
Corollary 5.4.8. Let f be a Lipschitz continuously differentiable function on Ω. Then if
x∗ is stationary then F (x∗ ) = 0.
5.5 Superlinear Convergence

Once the gradient projection iteration has identified the active constraints, PA(x∗ ) x∗ is known.
At that point the minimization problem for PI x∗ is unconstrained and, in principal, any super-
linearly convergent method for unconstrained optimization could then be used.
The problem with this idea is, of course, that determining when the active set has been
identified is possible only after the problem has been solved and an error in estimating the active
set can have devastating effects on convergence. In this section we discuss two approaches: one,
based on Newton’s method, is presented only as a local method; the other is a BFGS–Armijo
method similar to Algorithm bfgsopt.
We will begin with the development of the local theory for the projected Newton method
[19]. This analysis illustrates the important problem of estimation of the active set. As with the
unconstrained minimization problem, the possibility of negative curvature makes this method
difficult to globalize (but see §5.6 for pointers to the literature on trust region methods). Following
the approach in §4.2 we describe a projected BFGS–Armijo scheme in §5.5.3.
5.5.1 The Scaled Gradient Projection Algorithm

One might think that the theory developed in §5.4 applies equally well to iterations of the form
x+ = P(xc − λHc−1 ∇f (xc ))
where Hc is spd. This is not the case as the following simple example illustrates. Let N = 2,
Li = 0, and Ui = 1 for all i. Let
f (x) =
x − (−1, 1/2)T
2 /2;

then the only local minimizer for (5.3) is x∗ = (0, 1/2)T . Let xc = (0, 0) (not a local mini-
mizer!); then ∇f (xc ) = (1, −1/2)T . If

−1 2 1
Hc =
1 2
then Hc−1 (and hence Hc ) is spd, and

2 1 1 3/2
Hc−1 ∇f (xc ) = = .
1 2 −1/2 0
Therefore, for all λ > 0,

0 −1 1 −3λ/2 0
xc (λ) = P − λHc =P = = xc .
0 −1/2 0 0
The reason that xc (λ) = xc for all λ > 0 is that the search direction for the unconstrained
problem has been rotated by Hc−1 to be orthogonal to the direction of decrease in the inactive
directions for the constrained problem. Hence, unlike the constrained case, positivity of the
model Hessian is not sufficient and we must be able to estimate the active set and model the
reduced Hessian (rather than the Hessian) if we expect to improve convergence.
The solution proposed in [19] is to underestimate the inactive set in a careful way and
therefore maintain a useful spd approximate reduced Hessian. For x ∈ Ω and
0 ≤ < min(Ui − Li )/2,
we define A1 (x), the -active set at x, by
(5.29) A1 (x) = {i | Ui − (x)i ≤ or (x)i − Li ≤ }.
And let I 1 (x), the -inactive set, be the complement of A1 (x).

Given 0 ≤ c < min(Ui − Li )/2, xc , and an spd matrix Hc , we model the reduced Hessian
with Rc , the matrix with entries

 δij if i ∈ A1c (xc ) or j ∈ A1c (xc ),
Rc = PAc (xc ) + PI c (xc )Hc PI c (xc ) =

(Hc )ij otherwise.
(5.30)
When the explicit dependence on xc , c , and Hc is important we will write
R(xc , c , Hc ).
So, for example,

∇2R f (xc ) = R(xc , 0, ∇2 f (xc )).
Given 0 < < min(Ui − Li )/2 and an spd H, define
xH,1 (λ) = P(x − λR(x, , H)−1 ∇f (x)).
It requires proof that

f (xH,1 (λ)) < f (x)
for λ sufficiently small. We prove more and show that the sufficient decrease condition
(5.31) f (xH,1 (λ)) − f (x) ≤ −α∇f (x)T (x − xH,1 (λ))

holds for sufficiently small λ.

Lemma 5.5.1. Let x ∈ Ω, 0 < < min(Ui − Li )/2, and H be spd with smallest and largest
eigenvalues 0 < λs ≤ λl . Let ∇f be Lipschitz continuous on Ω with Lipschitz constant L. Then
there is λ̄(, H) such that (5.31) holds for all
(5.32) λ ≤ λ̄(, H).
Proof. The proof will show that
∇f (x)T (x − xH,1 (λ)) ≥ λ−1 T

l ∇f (x) (x − x(λ))
and then use the method of proof from Theorem 5.4.5. We do this by writing
∇f (x)T (x − xH,1 (λ)) = (PA (x) ∇f (x))T (x − xH,1 (λ)) + (PI (x) ∇f (x))T (x − xH,1 (λ))
and considering the two terms on the right side separately.

We begin by looking at (PA (x) ∇f (x))T (x − xH,1 (λ)). Note that
(xH,1 (λ))i = (x(λ))i for i ∈ A1 (x)
and, therefore,
(5.33) (PA (x) ∇f (x))T (x − xH,1 (λ)) = (PA (x) ∇f (x))T (x − x(λ)).
We will need to show that
(5.34) (PA (x) ∇f (x))T (x − x(λ)) ≥ 0.
Now assume that

min(Ui − Li )
(5.35) λ < λ̄1 = .
2 maxx∈Ω
∇f (x)
∞
Since A(x) ⊂ A1 (x) we can investigate the contributions of A(x) and A1 (x) ∩ I(x) separately.
If i ∈ A(x) then (5.35) implies that either (x − x(λ))i = λ(∇f (x))i or (x − x(λ))i = 0. In
either case (x − x(λ))i (∇f (x))i ≥ 0. If i ∈ A1 (x) ∩ I(x) and (x − x(λ))i = λ(∇f (x))i , then
it must be the case that i ∈ A(x(λ)) and therefore we must still have (x − x(λ))i (∇f (x))i ≥ 0.
Hence (5.34) holds.
Now if i ∈ I 1 (x) then, by definition,
Li + ≤ (x)i ≤ Ui −
and, hence, if

(5.36) λ ≤ λ̄2 =
maxx∈Ω
R(x, , H)−1 ∇f (x)
∞
then i is in the inactive set for both xH,1 (λ) and x(λ). Therefore, if (5.36) holds then
(PI (x) ∇f (x))T (x − xH,1 (λ)) = λ(PI (x) ∇f (x))T H −1 PI (x) ∇f (x)
(5.37) ≥ λ−1
l λ
−1

PI (x) (x − x(λ))
2
= λ−1 T
l (PI (x) ∇f (x)) (x − x(λ)).

Hence, using Corollary 5.4.4, (5.33), (5.34), and (5.37), we obtain
∇f (x)T (x − xH,1 (λ)) = (PA (x) ∇f (x))T (x − xH,1 (λ))
+(PI (x) ∇f (x))T (x − xH,1 (λ))
≥ (PA (x) ∇f (x))T (x − x(λ)) + λ−1 T

l (PI (x) ∇f (x)) (x − x(λ))
min(1,λ−1 )
≥ min(1, λ−1 T
l )∇f (x) (x − x(λ)) ≥ λ
l

x − x(λ)
2 .
(5.38)
The remainder of the proof is almost identical to that for Theorem 5.4.5. The fundamental
theorem of calculus and the Lipschitz continuity assumption imply that
f (xH,1 (λ)) − f (x) ≤ −∇f (x)T (x − xH,1 (λ)) + L

x − xH,1 (λ)
2 .
We apply (5.38) and obtain
f (xH,1 (λ)) − f (x) ≤ −(1 − Lλ max(1, λl ))∇f (x)T (x − xH,1 (λ)),
which implies (5.31) if 1 − Lλ max(1, λl ) ≥ α which will follow from
(1 − α)
(5.39) λ ≤ λ̄3 = .
max(1, λl )L
This completes the proof with λ̄ = min(λ̄1 , λ̄2 , λ̄3 ).

An algorithm based on these ideas is the scaled gradient projection algorithm. The name
comes from the scaling matrix H that is used to computed the direction. The inputs are the initial
iterate, the vectors of upper and lower bounds u and l, the relative-absolute residual tolerance
vector τ = (τr , τa ), and a limit on the number of iterations. Left unstated in the algorithmic
description are the manner in which the parameter is computed and the way in which the
approximate Hessians are constructed.
Algorithm 5.5.1. sgradproj(x, f, τ, nmax)
1. For n = 1, . . . , nmax
(a) Compute f and ∇f ; test for termination using (5.18).

(b) Compute and an spd H.
(c) Solve
R(x, , Hc )d = −∇f (xc ).
(d) Find the least integer m such that (5.13) holds for λ = β m .
(e) x = x(λ).
If our model reduced Hessians remain uniformly positive definite, a global convergence
result completely analogous to Theorem 3.2.4 holds.
Theorem 5.5.2. Let ∇f be Lipschitz continuous with Lipschitz constant L. Assume that the
matrices Hn are symmetric positive definite and that there are κ̄ and λl such that κ(Hn ) ≤ κ̄,
and
Hn
≤ λl for all n. Assume that there is ¯ > 0 such that ¯ ≤ n < min(Ui − Li )/2 for
all n.

Then
(5.40) lim
xn − xn (1)
= 0,
n→∞
and hence any limit point of the sequence of iterates produced by Algorithm sgradproj is a
stationary point.
In particular, if xnl → x∗ is any convergent subsequence of {xn }, then x∗ = x∗ (1). If xn
converges to a nondegenerate local minimizer x∗ , then the active set of xn is the same as that of
x∗ after finitely many iterations.
Proof. With Lemma 5.5.1 and its proof in hand, the proof follows the outline of the proof of
Theorem 5.4.6. We invite the reader to work through it in exercise 5.8.3.
5.5.2 The Projected Newton Method

The requirement in the hypothesis of Theorem 5.5.2 that the sequence {n } be bounded away
from zero is used to guarantee that the steplengths λn are bounded away from zero. This is
needed because appears in the numerator in (5.36). However, once the active set has been
identified and one is near enough to a nondegenerate local minimizer for the reduced Hessians to
be spd, one is solving an unconstrained problem. Moreover, once near enough to that minimizer,
the convergence theory for Newton’s method will hold. Then one can, in principle, set n = 0
and the iteration will be q-quadratically convergent. In this section we discuss an approach from
[19] for making a transition from the globally convergent regime described in Theorem 5.5.2 to
the locally convergent setting where Newton’s method converges rapidly.
If the initial iterate x0 is sufficiently near a nondegenerate local minimizer x∗ and we take
Hn = ∇2R f (xn )
in Algorithm sgradproj, then the resulting projected Newton method will take full steps (i.e.,
λ = 1) and, if n is chosen with care, converge q-quadratically to x∗ .
A specific form of the recommendation from [19], which we use here, is
(5.41) n = min(
xn − xn (1)
, min(Ui − Li )/2).
Note that while xn is far from a stationary point and the reduced Hessian is spd, then n will be
bounded away from zero and Theorem 5.5.2 will be applicable. The convergence result is like
Theorem 2.3.3 for local convergence but makes the strong assumption that Hn is spd (valid near
x∗ , of course) in order to get a global result.
Algorithm projnewt is the formal description of the projected Newton algorithm. It is a
bit more than just a specific instance of Algorithm gradproj. Keep in mind that if the initial
iterate is far from x∗ and the reduced Hessian is not spd, then the line search (and hence the
entire iteration) may fail. The algorithm tests for this. This possibility of indefiniteness is the
weakness in any line search method that uses ∇2 f when far from the minimizer. The inputs to
Algorithm projnewt are the same as those for Algorithm gradproj. The algorithm exploits
the fact that
(5.42) R(x, , ∇2R f (x)) = R(x, , ∇2 f (x))
which follows from A(x) ⊂ A1 (x).
Algorithm 5.5.2. projnewt(x, f, τ, nmax)
1. For n = 1, . . . , nmax
(a) Compute f and ∇f ; test for termination using (5.18).

(b) Set =
x − x(1)
.

(c) Compute and factor R = R(x, , ∇2R f (x)). If R is not spd, terminate with a failure
message.
(d) Solve Rd = −∇f (xc ).
(e) Find the least integer m such that (5.13) holds for λ = β m .
(f) x = x(λ).
Theorem 5.5.3. Let x∗ be a nondegenerate local minimizer. Then if x0 is sufficiently near

to x∗ and A(x0 ) = A(x∗ ) then the projected Newton iteration, with n =
xn − xn (1)
, will
converge q-quadratically to x∗ .
Proof. Our assumption that the active set has been identified, i.e.,
A(xc ) = A(x+ ) = A(x∗ ),
implies that
PA(xc ) ec = PA(xc ) e+ = 0.
Hence, we need only estimate PI(xc ) e+ to prove the result.
Let
δ ∗ = min∗ (|(x)i − Ui |, |(x)i − Li |) > 0.
i∈I(x )
We reduce
e
if necessary so that

e
≤ δ ∗ /M,
where M is the constant in Theorem 5.4.2. We may then apply Theorem 5.4.2 to conclude that
both c < δ ∗ and
ec
< δ ∗ . Then any index i ∈ A1c (xc ) must also be in A(xc ) = A(x∗ ).
Hence
(5.43) A1c (xc ) = A(xc ) = A(x∗ ).
From this we have
(5.44) R(xc , c , ∇2R f (xc )) = ∇2R f (xc ).
Hence, for
ec
sufficiently small the projected Newton iteration is
x+ = P(xc − (∇2R f (xc ))−1 ∇f (xc )).
By the fundamental theorem of calculus,
(5.45) ∇f (xc ) = ∇f (x∗ ) + ∇2 f (xc )ec + E1 ,
where 1
E1 = (∇2 f (x∗ + tec ) − ∇2 f (xc ))ec dt
0
and hence
E1
≤ K1
ec
2 for some K1 > 0.
By the necessary conditions,
(5.46) PI(x) ∇f (x∗ ) = PI(x∗ ) ∇f (x∗ ) = 0.
By the fact that I(xc ) = I(x∗ ), we have the equivalent statements
(5.47) ec = PI(xc ) ec and PA(xc ) ec = 0.

Therefore, combining (5.45), (5.46), (5.47),
PI(xc ) ∇f (xc ) = PI(xc ) ∇2 f (xc )PI(xc ) ec + PI(xc ) E1
(5.48) = PA(xc ) ec + PI(xc ) ∇2 f (xc )PI(xc ) ec + PI(xc ) E1
= ∇2R f (xc )ec + PI(xc ) E1 .
So, by definition of ∇2R ,
PI(xc ) (∇2R f (xc ))−1 ∇f (xc ) = (∇2R f (xc ))−1 PI(xc ) ∇f (xc ) = ec + E2 ,
where
E2
≤ K2
ec
2 for some K2 > 0.
Since PI(xc ) Pw = PPI(xc ) w for all w ∈ RN ,
PI(xc ) x+ = PI(xc ) P(xc − (∇2R f (xc ))−1 ∇f (xc ))
= PPI(xc ) (xc − (∇2R f (xc ))−1 ∇f (xc )) = P(x∗ − E2 ).
Therefore,
e+
≤ K2
ec
2 as asserted.
5.5.3 A Projected BFGS–Armijo Algorithm

We can apply the structured quasi-Newton updating scheme from §4.3 with
(5.49) C(x) = PA(x)
and update an approximation to the part of the model Hessian that acts on the inactive set. In
this way we can hope to maintain a positive definite model reduced Hessian with, say, a BFGS
update. So if our model reduced Hessian is
R = C(x) + A,
we can use (4.42) to update A (with A0 = PI 0 (x0 ) , for example), as long as the active set does
not change. If one begins the iteration near a nondegenerate local minimizer with an accurate
approximation to the Hessian, then one would expect, based on Theorem 5.5.3, that the active
set would remain constant and that the iteration would converge q-superlinearly.
However, if the initial data is far from a local minimizer, the active set can change with each
iteration and the update must be designed to account for this. One way to do this is to use a
projected form of the BFGS update of A from (4.42),
T
y# y# (Ac s)(Ac s)T
(5.50) A+ = PI+ Ac PI+ + − PI+ PI+ ,
y #T s sT Ac s
with
y # = PI+ (∇f (x+ ) − ∇f (xc )).
Here I+ = I 1+ (x+ ). This update carries as much information as possible from the previous
model reduced Hessian while taking care about proper approximation of the active set. As in
T
the unconstrained case, if y # s ≤ 0 we can either skip the update or reinitialize A to PI .
A is not spd if any constraints are active. However, we can demand that A be symmetric
indefinite, and a generalized inverse A† exists. We have
(5.51) (PA + A)−1 = PA + A† .

If A(xc ) = A(x+ ) any of the low-storage methods from Chapter 4 can be used to update A† .
In this case we have
T

† sy # † y # sT ssT
(5.52) A+ = I − T
A c I − T
+ T
.
y# s y# s y# s
Since s = PI+ s if A(x+ ) = A(xc ), we could replace s by s# = PI+ s in (5.52).

If the set of active constraints changes, then (5.52) is no longer true and we cannot replace
s by s# . One approach is to store A, update it as a full matrix, and refactor with each nonlinear
iteration. This is very costly for even moderately large problems. Another approach, used in
[250], is to reinitialize A = PI whenever the active set changes. The problem with this is that
in the terminal phase of the iteration, when most of the active set has been identified, too much
information is lost when A is reinitialized.
In this book we suggest an approach based on the recursive BFGS update that does not
discard information corresponding to that part of the inactive set that is not changed. The idea is
that even if the active set has changed, we can still maintain an approximate generalized inverse
with T
T
T
† s# y # † y # s# s# s#
(5.53) A+ = I − T
P A P
I+ c I+ I − T
+ T
.
y # s# y # s# y # s#
The formulation we use in the MATLAB code bfgsbound is based on (5.52) and Al-
gorithm bfgsrec. Algorithm bfgsrecb stores the sequences {yk# } and {s# k } and uses
Algorithm bfgsrec and the projection PI+ to update A† as the iteration progresses. The data
are the same as for Algorithm bfgsrec with the addition of
PIn = PI n (xn ) .
Note that the sequences {yk# } and {s#

k } are changed early in the call and then the unconstrained
algorithm bfgsrec is used to do most of the work.
Algorithm 5.5.3. bfgsrecb(n, {s# # †
k }, {yk }, A0 , d, PIn )
1. d = PIn d.
2. If n = 0, d = A†0 d; return
T # T #
3. α = s# n−1 d/yn−1 s# ; d = d − αyn−1
4. call bfgsrec(n − 1, {s# k }, {yk# }, A†0 , d)

# # T T
5. d = d + (α − (yn−1 d/yn−1 s# n−1 ))s# n−1
6. d = PIn d.
The projected BFGS–Armijo algorithm that we used in the example problems in §5.7 is
based on Algorithm bfgsrecb. Note that we reinitialize ns to zero (i.e., reinitialize A to PI )
#T
when yns s ≤ 0. We found experimentally that this was better than skipping the update.
Algorithm 5.5.4. bfgsoptb(x, f, τ, u, l)
1. ns = n = 0; pg0 = pg = x − P(x − ∇f (x))
2. = min(min(Ui − Li )/2,
pg
); A = A1 (x); I = I 1 (x); A0 = PI
3. While
pg
≤ τa + τr
pg0

(a) d = −∇f (x); Call bfgsrecb(ns, {s# # †

k }, {yk }, A0 , d, PI )
(b) d = −PA ∇f (x) + d
(c) Find the least integer m such that (5.13) holds for λ = β m . Set s#
ns = PI (x(λ) − x)
#
(d) xp = x(λ); y = ∇f (xp) − ∇f (x); x=xp; yns = PI (∇f (xp) − ∇f (x))
# T
(e) If yns s#
n s > 0 then ns = ns + 1, else ns = 0
(f) x = xp; pg = x − P(x − ∇f (x))
(g) = min(min(Ui − Li )/2,
pg
); A = A1 (x); I = I 1 (x)
(h) n = n + 1
Theorem 4.1.3 can be applied directly once the active set has been identified and a good
initial approximation to the reduced Hessian is available. The reader is invited to construct the
(easy!) proof in exercise 5.8.6.
Theorem 5.5.4. Let x∗ be a nondegenerate local minimizer. Then if x0 is sufficiently near
to x∗ , A(x0 ) = A(x∗ ), and A0 sufficiently near to PI(x∗ ) ∇2 f (x∗ )PI(x∗ ) , then the projected
BFGS iteration, with n =
xn − xn (1)
, will converge q-superlinearly to x∗ .
A global convergence result for this projected BFGS algorithm can be derived by combining
Theorems 5.5.2 and 4.1.9.
Theorem 5.5.5. Let ∇f be Lipschitz continuous on Ω. Assume that the matrices Hn are
constructed with the projected BFGS method (5.50) and satisfy the assumptions of Theorem 5.5.2.
Then (5.40) and the conclusions of Theorem 5.5.2 hold.
Moreover, if x∗ is a nondegenerate local minimizer such that there is n0 such that A(xn ) =
A(x∗ ) for all n ≥ n0 , Hn0 is spd, and the set
D = {x | f (x) ≤ f (xn0 ) and A(x) = A(x∗ )}
is convex, then the projected BFGS–Armijo algorithm converges q-superlinearly to x∗ .
5.6 Other Approaches

Our simple projected-BFGS method is effective for small to medium sized problems and for very
large problems that are discretizations of infinite-dimensional problems that have the appropriate
compactness properties. The example in §4.4.2 nicely illustrates this point. For other kinds of
large problems, however, more elaborate methods are needed, and we present some pointers to
the literature in this section.
The limited memory BFGS method for unconstrained problems described in [44] and [176]
has also been extended to bound constrained problems [42], [280]. More general work on line
search methods for bound constrained problems can be found in [47], [194], and [42].
Very general theories have been developed for convergence of trust region methods for
bound constrained problems. The notion of Cauchy decrease can, for example, be replaced by
the decrease from a gradient projection step for the quadratic model [191], [259], [66]. One
could look for minima of the quadratic model along the projection path [63], [64], or attempt to
project the solution of an unconstrained model using the reduced Hessian [162].
A completely different approach can be based on interior point methods. This is an active
research area and the algorithms are not, at least at this moment, easy to implement or analyze.
This line of research began with [57] and [58]. We refer the reader to [86] and [266] for more
recent accounts of this aspect of the field and to [140] and [79] for some applications to control
problems and an account of the difficulties in infinite dimensions.

Gradient Projection Gradient Projection

1 2
10 10
Projected Gradient Norm
Function Value
0
10
1 1
10 10
0 500 1000 1500 2000 0 500 1000 1500 2000
Projected BFGS Projected BFGS

5 2
10 10
Function Value
0
10
5
10
10 1
10 10
0 10 20 30 40 0 10 20 30 40
Figure 5.1: Solution to Constrained Parameter ID Problem
Solution to bound constrained control problem

1.4
1.3
1.2
1.1
0.9
0.8
0.7
0.6
0.5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Figure 5.2: Solution to Discrete Control Problem: First Example

5.6.1 Infinite-Dimensional Problems

The results in this part of this book do not extend in a direct way to infinite-dimensional problems.
One reason for this is that often infinite-dimensional problems have countably infinitely many
constraints or even a continuum of constraints; hence Ω is not compact in the norm topology of
the Banach space in which the problem is posed and appeals to various types of weak continuity
must be made (see [122] for an example of such arguments and [122] and [10] for applications).
Moreover, identification of an active set in finitely many iterations is not always possible. A
more complete account of this issue may be found in [254], [162], [161].
These are not the only complications that can arise in infinite dimension. Even the projected
gradient method presents challenges, especially if the minima fail to be nondegenerate in the
sense of this book [94], [95]. Convergence behavior for discretized problems can be different
from that for the continuous problem [97]. Nonequivalence of norms makes convergence results
difficult to formulate and analyze for both line search [96], [254], [98] and trust region [140],
[162] methods.
The functional analytic structure of many control problems can be exploited with fast mul-
tilevel methods. Both second kind multigrid methods from [138] and variants of the Atkinson–
Brakhage method [9], [31] have been applied to fixed point formulations of parabolic boundary
control problems in one space dimension [136], [137], [153], [162], [161].
5.7 Examples
The computations in this section were done with the MATLAB code bfgsbound. In this code
the storage is limited to five pairs of vectors, and β = .1 was used in the line search.
5.7.1 Parameter ID Problem

We consider the parameter problem from §3.4.1 with bounds L = (2, 0)T and U = (20, 5)T .
The initial iterate x0 = (5, 5)T is feasible, but the global minimum of (1, 1)T is not. As one
might expect, the lower bound constraint on (x)1 is active at the optimal point x∗ ≈ (2, 1.72)T .
The termination criterion for both the gradient projection and projected BFGS algorithms was

u − u(1)
≤ 10−6 .
The gradient projection algorithm failed. While the value of the objective function was
correct, the projected gradient norm failed to converge and the active set was not identified.
The projected BFGS iteration converged in 35 iterations. One can see the local superlinear
convergence in Figure 5.1 from the plot of the projected gradient norms. The cost of the BFGS
iteration was 121 function evaluations, 36 gradients, and roughly 5.3 million floating point
operations.

We base the two control problem examples on the example from §1.6.1.
Our first example takes N = 2000, T = 1, y0 = 0,
L(y, u, t) = (y − 3)2 + .1 ∗ u2 , and φ(y, u, t) = uy + t2 ,
with the bound constraints

.5 ≤ u ≤ 2,
and the initial iterate u0 = 2. We terminated the iteration when
u − u(1)
≤ 10−5 . In
Figure 5.2 we plot the solution of this problem. Clearly the active set is not empty for the
constrained problem.

Projected BFGS 4 Projected BFGS

2 x 10
10 1.74
0 1.73
10
Function Value
1.72
2
10
1.71
4
10
1.7
6
10 1.69
0 10 20 30 40 0 10 20 30 40
Gradient Projection 4 Gradient Projection

2 x 10
10 1.74
0 1.73
10
Function Value
1.72
2
10
1.71
4
10
1.7
6
10 1.69
0 50 100 0 50 100
Figure 5.3: Constrained Discrete Control Problem I
Projected BFGS Projected BFGS

5 8
10 10
Function Value
0 6
10 10
5 4
10 10
10 2
10 10
0 2 4 6 0 2 4 6
Gradient Projection Gradient Projection

5 8
10 10
Function Value
0 6
10 10
5 4
10 10
10 2
10 10
0 2 4 6 8 0 2 4 6 8
Figure 5.4: Constrained Discrete Control Problem II

We solve the constrained problem with Algorithm gradproj and Algorithm bfgsoptb.
In Figure 5.3 we plot the function value and the norm of the projected gradient u − u(1).
The projected BFGS iteration required 71 function evaluations, 36 gradient evaluations, and
roughly 5.6 million floating point operations, while the gradient projection needed 183 function
evaluations, 92 gradient evaluations, and roughly 10.4 million floating point operations.
Our second control problem example solves the same problem as in §3.4.2 using the con-
straints
−206 ≤ u ≤ 206.
We terminate the iteration when
u − u(1)
≤ 10−6 , which is exactly the condition used in
§3.4.2 when the active set is empty. The solution to the unconstrained problem is feasible,
the active set is empty, and the initial iterate is feasible. Both the gradient projection iteration
and the projected BFGS iteration converge to the solution of the unconstrained problem. The
constraints are not active at either the initial iterate or the final solution but are active inside
the line search for the first iterate and for the second iterate. As is clear from a comparison
of Figures 5.4 and 3.3, this small change has a dramatic effect on the cost of the optimization,
eliminating the need for the scaling fixup (3.50). The gradient projection method, requiring 15
function evaluations, 8 gradient evaluations, and roughly 167 thousand floating point operations,
is far more efficient that the steepest descent iteration reported in §3.4.2. The projected BFGS
iteration was somewhat worse, needing 223 thousand operations, but only 13 function evaluations
and 7 gradient evaluations. In this example the cost of maintaining the BFGS update was not
compensated by a significantly reduced iteration count.
5.8 Exercises on Bound Constrained Optimization

5.8.1. Suppose that f is continuously differentiable, that x∗ is a nondegenerate local minimizer
for problem (5.4), and all constraints are active. Show that there is δ such that
1. if x ∈ B(δ) then x∗ = P(x − ∇f (x)), and
2. the gradient projection algorithm converges in one iteration if x0 ∈ B(δ).
5.8.2. Show that if H = I then (5.31) and (5.13) are equivalent.
5.8.4. Verify (5.42).
5.8.5. Suppose the unconstrained problem (1.2) has a solution x∗ at which the standard assump-
tions for unconstrained optimization hold. Consider the bound constrained problem (5.3)
for u and l such that x∗ ∈ Ω and A(x∗ ) is not empty. Is x∗ a nondegenerate local mini-
mizer? If not, how are the results in this chapter changed? You might try a computational
example to see what’s going on.
5.8.7. Verify (5.51).
5.8.8. Verify (5.52).
5.8.9. Formulate a generalization of (4.33) for updating A† .
5.8.10. What would happen in the examples if we increased the number of (y, s) pairs that were
stored? By how much would the BFGS cost be increased?

Part II
Optimization of Noisy Functions


Chapter 6
Basic Concepts and Goals
The algorithms in Part I cannot be implemented at all if the gradient of f is not available,
either analytically or via a difference. Even if gradients are available, these algorithms are
not satisfactory if f has many local minima that are not of interest. We limit our coverage to
deterministic sampling algorithms which are generally applicable and are more or less easy to
implement. Of these algorithms, only the DIRECT algorithm [150] covered in §8.4.2 is truly
intended to be a global optimizer.
The study of optimization methods that do not require gradients is an active research area (see
[227] for a survey of some of this activity), even for smooth problems [61], [62]. Even though
some of the methods, such as the Nelder–Mead [204] and Hooke–Jeeves [145] algorithms are
classic, most of the convergence analysis in this part of the book was done after 1990.
The algorithms and theoretical results that we present in this part of the book are for objective
functions that are perturbations of simple, smooth functions. The surfaces in Figure 6.1 illustrate
this problem. The optimization landscape on the left of Figure 6.1, taken from [271], arose in
a problem in semiconductor design. The landscape on the right is a simple perturbation of a
convex quadratic.
4.5
3.5
20
3
0
2.5
-20
2
-40
1.5
-60
25
1
20
-80
0 15
5 0.5
10
10
15 5
20 0
25 0 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
Figure 6.1: Optimization Landscapes
We do not discuss algorithms that explicitly smooth the objective function or apply a filter,
such as the ones in [168] and [187]. For general problems, these must sample the variable
space in some way, for example by performing high-dimensional integration, and are too costly.
However, in some special cases these integrals can be performed analytically and impressive
results for special-purpose filtering algorithms for computational chemistry have been reported
in, for example, [196] and [277]. Nor do we discuss analog methods (see [149] for a well-known
111
example).
We also omit stochastic methods like the special-purpose methods discussed in [38] and [39],
or more radical general-purpose global optimization algorithms, such as simulated annealing
[166] (see [1] and [265] for surveys of recent work), interval methods [152], or genetic algorithms
[143], [144] (see [246] or [123] for a survey), which are random to some extent or random
search algorithms. These probabilistic methods, however, should be considered when the more
conservative algorithms such as the ones in this part of the book fail.
6.1 Problem Statement

Consider an objective function f that is a perturbation of a smooth function fs by a small function
φ
(6.1) f (x) = fs (x) + φ(x).
Small oscillations in φ could cause f to have several local minima that would trap any conven-
tional gradient-based algorithms. The perturbation φ can, in general, be random or based on the
output of an experiment, [250], and may not return the same value when called twice with the
same argument. Hence φ need not even be a function. We assume that φ is everywhere defined
and bounded to make the statement of the results simpler.
6.2 The Simplex Gradient

Most of the the algorithms in this part of the book examine a simplex of points in RN at each
iteration and then change the simplex in response. In this section we develop the tools needed to
describe and analyze these algorithms. The fundamental idea is that many sampling algorithms
require enough information to approximate the gradient by differences and that the accuracy in
that difference approximation can be used to analyze the convergence. However, for problems
of the form (6.1), one must take care not to make the difference increments so small as to attempt
to differentiate the noise.
The ideas in this section were originally used in [155] to analyze the Nelder–Mead [204]
algorithm, which we discuss in §8.1. However, the ideas can be applied to several classes of
algorithms, and we follow the development in [29] in this section.
Definition 6.2.1. A simplex S in RN is the convex hull of N + 1 points, {xj }N +1
j=1 . xj is
the jth vertex of S. We let V (or V (S)) denote the N × N matrix of simplex directions
V (S) = (x2 − x1 , x3 − x1 , . . . , xN +1 − x1 ) = (v1 , . . . , vN ).
We say S is nonsingular if V is nonsingular. The simplex diameter diam(S) is
diam(S) = max
xi − xj
.
1≤i,j≤N +1
We will refer to the l2 condition number κ(V ) of V as the simplex condition.

We let δ(f : S) denote the vector of objective function differences
δ(f : S) = (f (x2 ) − f (x1 ), f (x3 ) − f (x1 ), . . . , f (xN +1 ) − f (x1 ))T .
We will not use the simplex diameter directly in our estimates or algorithms. Rather we will use
two oriented lengths
σ+ (S) = max
x1 − xj
and σ− (S) = min
x1 − xj
.
2≤j≤N +1 2≤j≤N +1
Clearly,
σ+ (S) ≤ diam(S) ≤ 2σ+ (S).

BASIC CONCEPTS AND GOALS 113
6.2.1 Forward Difference Simplex Gradient
Definition 6.2.2. Let S be a nonsingular simplex with vertices {xj }N +1

j=1 . The simplex gradient
D(f : S) is
D(f : S) = V −T δ(f : S).
Note that the matrix of simplex directions and the vector of objective function differences
depend on which of the vertices is labeled x1 . Most of the algorithms we consider in this part
of the book use a vertex ordering or sample on a regular stencil. In this way the algorithms, in
one way or another, use a simplex gradient.
This definition of simplex gradient is motivated by the first-order estimate in Lemma 6.2.1.
Lemma 6.2.1. Let S be a simplex. Let ∇f be Lipschitz continuous in a neighborhood of S
with Lipschitz constant 2Kf . Then there is K > 0, depending only on Kf such that
(6.2)
∇f (x1 ) − D(f : S)
≤ Kκ(V )σ+ (S).
Proof. Our smoothness assumptions on f and Taylor’s theorem imply that for all 2 ≤ j ≤
N + 1,
|f (x1 ) − f (xj ) + vjT ∇f (x1 )| ≤ Kf
vj
2 ≤ Kf σ+ (S)2 .
Hence

δ(f : S) − V T ∇f (x1 )
≤ N 1/2 Kf σ+ (S)2
and hence, setting K = N 1/2 Kf ,

∇f (x1 ) − D(f : S)
≤ K
V −T
σ+ (S)2 .
The conclusion follows from the fact that σ+ (S) ≤

V
.
Search algorithms are not intended, of course, for smooth problems. Minimization of ob-
jective functions of the form in (6.1) is one of the applications of these methods. A first-order
estimate that takes perturbations into account is our next result.
We will need to measure the perturbations on each simplex. To that end we define for any
set T

φ
T = sup
φ(x)
.
x∈T
A first-order estimate also holds for the simplex gradient of an objective function that satisfies
(6.1).
Lemma 6.2.2. Let S be a nonsingular simplex. Let f satisfy (6.1) and let ∇fs be Lipschitz
continuous in a neighborhood of S with Lipschitz constant 2Ks . Then there is K > 0, depending
only on Ks , such that

φ
S
(6.3)
∇fs (x1 ) − D(f : S)
≤ Kκ(V ) σ+ (S) + .
σ+ (S)
Proof. Lemma 6.2.1 (applied to fs ) implies

∇fs (x1 ) − D(fs : S)
≤ Ks N 1/2 κ(V )σ+ (S).


√
Now, since
δ(φ : S)
≤ 2 N
φ
S , and σ+ (S) ≤
V
,

D(f : S) − D(fs : S)
≤
V −T
δ(f : S) − δ(fs : S)
=
V −T
δ(φ : S)

φ
S
≤ 2N 1/2
V −T
φ
S ≤ 2N 1/2 κ(V ) .
σ+ (S)
This completes the proof with K = N 1/2 Ks + 2N 1/2 .

The constants K in (6.2) and (6.3) depend on S only through the Lipschitz constants of fs
and ∇fs in a neighborhood of S. We will express that dependence as K = K(S) when needed.
The algorithms in this section are most profitably applied to problems of the form (6.1), and
the goal is to extract as much information as possible from the smooth part fs of f without
wasting effort in a futile attempt to minimize the noise. In order to formulate our goal for
convergence clearly, we explore the consequences of a small simplex gradient in the special (and
not uncommon) case that the amplitude of the noise is small in Lemma 6.2.3.
Lemma 6.2.3. Let f satisfy (6.1) and let ∇fs be continuously differentiable in a compact set
Ω ⊂ RN . Assume that fs has a unique critical point x∗ in Ω. Then there is KΩ > 0 such that
for any simplex S ⊂ Ω with vertices {xj }N +1
j=1 ,

φ
S

x1 − x∗
≤ KΩ
D(f : S)
+ κ(V ) σ+ (S) + .
σ+ (S)
Proof. The compactness of Ω and our smoothness assumptions on fs imply that there is β0
such that

∇fs (x)
≥ β0
x − x∗
for all x ∈ Ω. We apply (6.3) to obtain

x1 − x∗
≤ β0−1
∇fs (x1 )

φ
S
≤ β0−1
D(f : S)
+ Kκ(V ) σ+ (S) + .
σ+ (S)
This completes the proof with KΩ = β0−1 max(1, K).

By sampling in an organized way simplex-based algorithms, some directly and some implic-
itly, attempt to drive the simplex gradient to a small value by changing the size of the simplices
over which f is sampled. The motion of the simplices and the scheme for changing the size
(especially the reduction in size) accounts for the differences in the algorithms. Theorem 6.2.4,
a direct consequence of Lemma 6.2.3, quantifies this. We will consider a sequence of uniformly
well-conditioned simplices. Such simplices are generated by several of the algorithms we will
study later.
Theorem 6.2.4. Let f satisfy (6.1) and let ∇fs be continuously differentiable in a compact
set Ω ⊂ RN . Assume that fs has a unique critical point x∗ in Ω. Let S k be a sequence of
simplices having vertices {xkj }N +1
j=1 . Assume that there is M such that
S k ⊂ Ω and κ(V (S k )) ≤ M for all k.
Then,
1. if

φ
S k
lim σ+ (S k ) = 0, lim = 0,
k→∞ k→∞ σ+ (S k )

and lim supk→∞

D(f : S k )
= , for some > 0, then there is KS > 0 such that
lim sup
x∗ − xk1
≤ KS ;
k→∞
2. if, for some > 0,
lim sup
φ
S k ≤ 2 , lim inf σ+ (S k ) ≥ , and lim inf
D(f : S k )
≤ ,
k→∞ k→∞ k→∞
then there is KS > 0 such that
lim sup
x∗ − xk1
≤ KS ( + lim sup σ+ (S k )).
k→∞ k→∞
6.2.2 Centered Difference Simplex Gradient

In this section we define the centered difference simplex gradient and prove a second-order
estimate. We will then prove two variants of Theorem 6.2.4, one to show how the role of the
noise φ differs from that in the one-sided derivative case and a second to quantify how the values
of f on the stencil can be used to terminate an iteration.
Definition 6.2.3. Let S be a nonsingular simplex in RN with vertices {xj }N +1
j=1 and simplex
directions vj = xj+1 − x1 . The reflected simplex R = R(S) is the simplex with vertices x1 and
rj = x1 − vj for j = 1, . . . , N.
The central simplex gradient DC (f : S) is
D(f : S) + D(f : R) V −T (δ(f : S) − δ(f : R))

DC (f : S) = = .
2 2
For example, if N = 1 and x2 = x1 + h, then r2 = x1 − h. Hence
f (x1 + h) − f (x1 ) f (x1 − h) − f (x1 )

D(f : S) = and D(f : R) = .
h −h
Therefore,
f (x1 + h) − f (x1 − h)
DC (f : S) = DC (f : R) =
2h
is the usual central difference.
Lemmas 6.2.5 and 6.2.6 are the second-order analogues of Lemmas 6.2.1 and 6.2.2.
Lemma 6.2.5. Let S be a nonsingular simplex and let ∇2 f be Lipschitz continuous in a
neighborhood of S ∪ R(S) with Lipschitz constant 3KC . Then there is K > 0 such that
(6.4)
∇f (x1 ) − DC (f : S)
≤ Kκ(V )σ+ (S)2 .
Proof. The Lipschitz continuity assumption implies that for all 2 ≤ j ≤ N + 1,

f (xj ) − f (rj ) + 2∇f (x1 )T vj ≤ KC
vj
3 ≤ Kc σ+ (S)3 .

As in the proof of Lemma 6.2.1 we have

V T (δ(f : S) − δ(f : R)) − V T ∇f (x1 )
≤ N 1/2 KC σ+ (S)3 ,
and hence the result follows with K = N 1/2 KC .

Lemma 6.2.6. Let S be a nonsingular simplex. Let f satisfy (6.1) and let ∇2 fs be Lipschitz
continuous in a neighborhood of S ∪ R(S) with Lipschitz constant 3KCs . Then there is K > 0,
depending only on KCs , such that

φ
S
(6.5)
∇fs (x1 ) − DC (f : S)
≤ Kκ(V ) σ+ (S)2 + .
σ+ (S)
Proof. This proof is very similar to that of Lemma 6.2.2 and is left to the reader.
The quality of the information that can be obtained from the central simplex gradient is
higher than that of the forward. The difference in practice can be dramatic, as the examples
in §7.6 illustrate. The consequences of a small central simplex gradient follow directly from
Lemma 6.2.6.
Lemma 6.2.7. Let f satisfy (6.1) and let ∇2 fs be continuously differentiable in a compact
set Ω ⊂ RN . Assume that fs has a unique critical point x∗ in Ω. Then there is KΩ > 0 such
that if a simplex S and its reflection R(S) are both contained in Ω then

φ
S

x1 − x∗
≤ KΩ
DC (f : S)
+ κ(V ) σ+ (S)2 + .
σ+ (S)
Lemma 6.2.7 is all one needs to conclude convergence from a sequence of small central
simplex gradients.
Theorem 6.2.8. Let f satisfy (6.1) and let ∇2 fs be continuously differentiable in a compact
set Ω ⊂ RN . Assume that fs has a unique critical point x∗ in Ω. Let S k be a sequence of
simplices having vertices {xkj }N +1
j=1 . Assume that there is M such that
S k , R(S k ) ⊂ Ω and κ(V (S k )) ≤ M for all k.
Then,
1. if

φ
S k
lim σ+ (S k ) = 0, lim = 0,
k→∞ k→∞ σ+ (S k )
and lim supk→∞
DC (f : S k )
= , for some > 0, then there is KS > 0 such that
lim sup
x∗ − xk1
≤ KS ;
k→∞
2. if, for some > 0,
lim sup
φ
S k ≤ 3 , lim inf σ+ (S k ) ≥ 2 , and lim inf
DC (f : S k )
≤ 2 ,
k→∞ k→∞ k→∞
then there is KS > 0 such that
lim sup
x∗ − xk1
≤ KS ( + lim sup σ+ (S k ))2 .
k→∞ k→∞

Theorem 6.2.8, like Theorem 6.2.4, motivates using a small simplex gradient as a test for
convergence. Suppose
φ
∞ ≤ and an algorithm generates sequences of simplices whose
vertices are intended to approximate a minimizer of fs . We can use the results in §2.3.1 to con-
clude that simplices with σ+ (S) << 1/2 will result in inaccurate forward difference gradients
and those with σ+ (S) << 2/3 in inaccurate central difference gradients. This indicates that
the central simplex gradient will be less sensitive to noise than the forward. While this is not
usually critical in computing a difference Hessian, where the loss of accuracy may cause slow
convergence, it can cause failure of the iteration if one is computing a difference gradient.
If one wants to terminate the algorithm when the simplex gradient is small, say, ≤ τ , a rough
estimate of the minimal possible value of τ is τ = O(1/2 ) for a forward difference simplex
gradient and τ = O(2/3 ) for a central simplex gradient.
Moreover, if one is using a centered difference, one has information on the values of f at
enough points to make an important qualitative judgment. In order to evaluate a central simplex
gradient f must be sampled at x1 and x1 ± vj for 1 ≤ j ≤ N . If f (x1 ) ≤ f (x1 ± vj ) for
all 1 ≤ j ≤ N , then one can question the validity of using the simplex gradient as a descent
direction or as a measure of stationarity. We call this stencil failure. We will use stencil failure
as a termination criterion in most of the algorithms we discuss in this part of the book. Our basis
for that is a result from [29], which only requires differentiability of fs .
Theorem 6.2.9. Let S be a nonsingular simplex such that for some µ− ∈ (0, 1) and κ+ > 0,
(6.6) κ(V ) ≤ κ+ and xT V V T x ≥ µ− σ+ (S)2

x
2 for all x.
Let f satisfy (6.1) and let ∇fs be Lipschitz continuously differentiable in a ball B of radius
2σ+ (S) about x1 . Assume that
(6.7) f (x1 ) < min{f (x1 ± vj )}.

j
Then, if K is the constant from Lemma 6.2.2,

φ
B
(6.8)
∇fs (x1 )
≤ 8µ−1
− Kκ + σ + (S) + .
σ+ (S)
Proof. Let R(S), the reflected simplex, have vertices x1 and {rj }N
j=1 . (6.7) implies that
each component of δ(f : S) and δ(f : R) is positive. Now since
V = V (S) = −V (R),
we must have
0 < δ(f : S)T δ(f : R)
(6.9) = (V T V −T δ(f : S))T (V (R)T V (R)−T δ(f : R))
= −D(f : S)T V V T D(f : R).

We apply Lemma 6.2.2 to both D(f : S) and D(f : R) to obtain
D(f : S) = ∇fs (x1 ) + E1 and D(f : R) = ∇fs (x1 ) + E2 ,
where, since κ(V ) = κ(V (R)) ≤ κ+ ,

φ
B

Ek
≤ Kκ+ σ+ (S) + .
σ+ (S)

Since
V
≤ 2σ+ (S) we have by (6.9)
∇fs (x1 )T V V T ∇fs (x1 ) ≤ 4σ+ (S)2

∇fs (x1 )
(
E1
+
E2
)
(6.10)
+4σ+ (S)2
E1
E2
.
The assumptions of the lemma give a lower estimate of the left side of (6.10),
wT V V T w ≥ µ− σ+ (S)2
w
2 .
Hence,

∇2 f (x1 )
≤ b
∇2 f (x1 )
+ c,
where, using (6.10),

φ
B
b= 8µ−1
1 K s κ+ σ+ (S) +
σ+ (S)
and 2

φ
B µ− 2
c= 4µ−1
− (Ks κ+ )
2
σ+ (S) + = B .
σ+ (S) 16
So b2 − 4c = b2 (1 − µ− /4) and the quadratic formula then implies that
√
2 b + b2 − 4c 1 + 1 − µ− /4

∇ f (x1 )
≤ =b ≤b
2 2
as asserted.
6.3 Examples
Our examples are selected to represent a variety of problems that can be attacked by the methods
in this part of the book and, at the same time, are easy for the reader to implement. Many of the
problems to which these methods have been applied have complex objective functions and have
been solved as team efforts [107], [250], [121], [70], [69]. In many such cases the objective
function is not even available as a single subroutine as the optimizer, simulator, and design tool
are one package. Hence, the examples we present in this part of the book are even more artificial
than the ones in the first part. The cost of an evaluation of f is much less in these examples than
it is in practice.
6.3.1 Weber’s Problem

Our discussion of this problem is based on [182]. Weber’s problem is to locate a central facility
(a warehouse or factory, for example) so that the total cost associated with distribution to several
demand centers is minimized. The model is that the cost is proportional to the distance from the
facility. The proportionality constant may be positive reflecting transportation costs or negative
reflecting environmental concerns.
If the locations of the demand centers are {zi } ⊂ R2 and the corresponding weights are
{wi }, then the objective function is

(6.11) f (x) = wi
x − zi
= wi [(x)1 − (zi )1 ]2 + [(x)2 − (zi )2 ]2 .
i i
We will assume that

wi > 0,
i


so that a global optimum exists. If i wi < 0 then inf f = −∞ and there is no global optimum.
Weber’s problem is not differentiable at x = zi because the square root function is not
differentiable at 0. A gradient-based algorithm, applied in a naive way, will have difficulty with
this problem. There are special-purpose algorithms (see [182] for a survey) for Weber’s problem,
especially if all the weights are positive. Our main interest is in the case where at least one weight
is negative. In that case there may be multiple local minima.
We will consider two examples. The first, and simplest, is from [182]. This example has
three demand centers with

2 90 43
w = (2, 4, −5)T and (z1 , z2 , z3 ) = .
42 11 88
The global minimum is at x∗ = (90, 11)T , at which the gradient is not defined. The complex
contours near the minimizer in Figure 6.2 illustrate the difficulty of the problem.
200
150
100
50
−50
−100
−150
−200
−200 −150 −100 −50 0 50 100 150 200
Figure 6.2: Contour/Surface for Weber’s Function: First Example
Our second example has two local minimizers, at (−10, −10) and (25, 30) with the global
minimizer at (25, 30). There are four demand centers with

T −10 0 5 25
w = (2, −4, 2, 1) and (z1 , z2 , z3 , z4 ) = .
−10 0 8 30
See Figure 6.3.
Our third example adds the oscillatory function
φ(x) = sin(.0035xT x) + 5 sin(.003(x − y)T (x − y))
to the second example, where y = (−20, 0)T . This complicates the optimization landscape
significantly, as the surface and contour plots in Figure 6.4 show.
6.3.2 Perturbed Convex Quadratics

The sum of a simple convex quadratic and low-amplitude high-frequency perturbation will serve
as a model problem for all the algorithms in this section. For example, the function graphed on
the right in Figure 6.1,
f (x) = 2x2 (1 + .75 cos(80x)/12) + cos(100x)2 /24

is one of the examples in [120]. Our general form will be
f (x) = (x − ξ0 )T H(x − ξ0 )(1 + a1 cos(bT1 (x − ξ1 ) + c1 (x − ξ1 )T (x − ξ1 )))
(6.12)
+a2 (1 + cos(bT2 (x − ξ2 )T + c2 (x − ξ2 )T (x − ξ2 ))) + a3 |rand|,

60
40
20
−20
−40
−60
−60 −40 −20 0 20 40 60
Figure 6.3: Contour/Surface for Weber’s Function: Second Example
60
40
20
−20
−40
−60
−60 −40 −20 0 20 40 60
Figure 6.4: Contour/Surface plots for Weber’s Function: Third Example
where {ξj }, {aj }, {bj }, {cj } are given and rand is a random number generator. f has been
designed so that the minimum value is O(a1 +a2 +a3 ). The unperturbed case a1 = a2 = a3 = 0
is also of interest for many of the algorithms in this part of the book.
6.3.3 Lennard–Jones Problem

The objective function is a simple model of the potential energy in a molecule of identical atoms.
Assume that there are M atoms and that ξi ∈ R3 is the position of the ith atom. Letting
dij =
ξi − ξj
and
T T
x = (ξ1T , . . . , ξM ) ∈ RN
where N = 3M , we have that the Lennard–Jones energy function is
M i−1
(6.13) f (x) = d−12
ij − 2d−6
ij .
i=1 j=1
2
f has many local minimizers (O(eM ) is one conjecture [142]) and the values at the mini-
mizers are close. Hence, the Lennard–Jones function does not conform to the noisy perturbation
of a smooth function paradigm. The reader is asked in some of the exercises to see how the
methods perform.

6.4 Exercises on Basic Concepts

6.4.1. Show that if wi > 0 for all i then Weber’s problem has a unique local minimum.
6.4.3. Try to minimize the Lennard–Jones functional using some of the algorithms from the first
part of the book. Vary the initial iterate and M . Compare your best results with those in
[142], [40], and [210].


Chapter 7
Implicit Filtering
7.1 Description and Analysis of Implicit Filtering

The implicit filtering algorithm was originally formulated in [270], [251], and [271], as a
difference-gradient implementation of the gradient projection algorithm [18] in which the dif-
ference increment is reduced in size as the iteration progresses. A different formulation for
unconstrained problems with certain convexity properties was introduced at about the same time
in [279]. From the point of view of this book, the simplex gradient is used in a direct way. The
algorithmic description and analysis in this chapter uses the results from §6.2 directly. We will
focus on unconstrained problems and derive the convergence results that implicit filtering shares
with the search algorithms in Chapter 8.
Implicit filtering, by using an approximate gradient directly, offers the possibility of im-
proved performance with quasi-Newton methods and can be easily applied to bound constrained
problems. We explore these two possibilities in §§7.2 and 7.4.
In its simplest unconstrained form, implicit filtering is the steepest descent algorithm with
difference gradients, where the difference increment varies as the iteration progresses. Because
the gradient is only an approximation, the computed steepest descent direction may fail to be a
descent direction and the line search may fail. In this event, the difference increment is reduced.
For a given x ∈ RN and h > 0 we let the simplex S(x, h) be the right simplex from x with
edges having length h. Hence the vertices are x and x + hvi for 1 ≤ i ≤ N with V = I. So
κ(V ) = 1. The performance of implicit filtering with a central difference gradient is far superior
to that with the forward difference gradient [120], [187], [250]. We will, therefore, use centered
differences in the discussion. We illustrate the performance of forward difference gradients in
§7.6.
We set
∇h f (x) = DC (f : S(x, h)).
We use a simple Armijo [7] line search and demand that the sufficient decrease condition
(7.1) f (x − λ∇h f (x)) − f (x) < −αλ

∇h f (x)
2
holds (compare with (3.4)) for some α > 0.

Our central difference steepest descent algorithm fdsteep terminates when
(7.2)
∇h f (x)
≤ τ h
for some τ > 0, when more than pmax iterations have been taken, after a stencil failure, or
when the line search fails by taking more than amax backtracks. Even the failures of fdsteep
123
can be used to advantage by triggering a reduction in h. The line search parameters α, β and
the parameter τ in the termination criterion (7.2) do not affect the convergence analysis that we
present here but can affect performance.
Algorithm 7.1.1. fdsteep(x, f, pmax, τ, h, amax)
1. For p = 1, . . . , pmax
(a) Compute f and ∇h f ; terminate if (6.7) or (7.2) hold.
(b) Find the least integer 0 ≤ m ≤ amax such that (7.1) holds for λ = β m . If no such
m exists, terminate.
(c) x = x − λ∇h f (x).
Algorithm fdsteep will terminate after finitely many iterations because of the limits on
the number of iterations and the number of backtracks. If the set {x | f (x) ≤ f (x0 )} is bounded
then the iterations will remain in that set. Implicit filtering calls fdsteep repeatedly, reducing
h after each termination of fdsteep. Aside from the data needed by fdsteep, one must
provide a sequence of difference increments, called scales in [120].
Algorithm 7.1.2. imfilter1(x, f, pmax, τ, {hk }, amax)
1. For k = 0, . . .
Call fdsteep(x, f, pmax, τ, hk , amax)
The convergence result follows from the second-order estimate, (6.5), the consequences of a
stencil failure, Theorem 6.2.9, and the equalities hk = σ+ (S k ) and κ(V k ) = 1. A similar result
for forward differences would follow from (6.3).
Theorem 7.1.1. Let f satisfy (6.1) and let ∇fs be Lipschitz continuous. Let hk → 0, {xk }
be the implicit filtering sequence, and S k = S(x, hk ). Assume that (7.1) holds (i.e., there is no
line search failure) for all but finitely many k. Then if
(7.3) lim (hk + h−1

k
φ
S k ) = 0
k→∞
then any limit point of the sequence {xk } is a critical point of fs .

Proof. If either (7.1) or (6.7) hold for all but finitely many k then, as is standard,
∇hk f (xk ) = DC (f : S k ) → 0.
Hence, using (7.3) and Lemma 6.2.2,
∇fs (xk ) → 0,
as asserted.
7.2 Quasi-Newton Methods and Implicit Filtering

The unique feature of implicit filtering is the possibility, for problems that are sufficiently smooth
near a minimizer, to obtain faster convergence in the terminal phase of the iteration by using a
quasi-Newton update of a model Hessian. This idea was first proposed in [250] and [120].
We begin with a quasi-Newton form of Algorithm fdsteep. In this algorithm a quasi-
Newton approximation to the Hessian is maintained and the line search is based on the quasi-
Newton direction
d = −H −1 ∇h f (x)

IMPLICIT FILTERING 125
terminating when either

(7.4) f (x + λd) − f (x) < αλ∇h f (x)T d
or too many stepsize reductions have been taken. With the application to implicit filtering in
mind, Algorithm fdquasi replaces the quasi-Newton H with the identity matrix when the line
search fails.
Algorithm 7.2.1. fdquasi(x, f, H, pmax, τ, h, amax)
1. For p = 1, . . . , pmax
(a) Compute f , ∇h f and d = −H −1 ∇h f ; terminate if (7.2) holds.
(b) Find the least integer 0 ≤ m ≤ amax such that (7.4) holds for λ = β m .
(c) x = x + λd.
(d) Update H with a quasi-Newton formula.
In the context of implicit filtering, where N is small, the full quasi-Newton Hessian or its
inverse is maintained throughout the iteration. Our MATLAB codes store the model Hessian.
Algorithm 7.2.2. imfilter2(x, f, pmax, τ, {hk }, amax)
1. H = I.
2. For k = 0, . . .
Call fdquasi(x, f, H, pmax, τ, hk , amax).
In [250] and [120] the SR1 method was used because it performed somewhat better than the
BFGS method in the context of a particular application. The examples in §7.6 show the opposite
effect, and both methods have been successfully used in practice.
7.3 Implementation Considerations

Implicit filtering has several iterative parameters and requires some algorithmic decisions in its
implementation. The parameters pmax, amax, and β play the same role that they do in any line
search algorithm. In our MATLAB code imfil.m, which we used for all the computations
reported in this book, we set pmax = 200 ∗ n, amax = 10, and β = 1/2.
The performance of implicit filtering can be sensitive to the value of τ [250], with small values
of τ leading to stagnation and values of τ that are too large leading to premature termination of
fdquasi. Using stencil failure as a termination criterion reduces the sensitivity to small values
of τ and we use τ = .01 in the computations.
The sequence of scales is at best a guess at the level of the noise in the problem. If several of
the scales are smaller than the level of the noise, the line search will fail immediately and work
at these scales will be wasted. Our implementation attempts to detect this by terminating the
optimization if the x is unchanged for three consecutive scales.
The simplex gradient may be a very poor approximation to the gradient. In some such cases
the function evaluation at a trial point may fail to return a value [250] and one must either trap
this failure and return an artificially large value, impose bound constraints, or impose a limit on
the size of the step. In our computations we take the latter approach and limit the stepsize to
10h by setting


 −H −1 ∇h f (x) if
H −1 ∇h f (x)
≤ 10h,

(7.5) d=

 −10hH −1 ∇h f (x)
 otherwise.

H −1 ∇h f (x)

The choice of a quasi-Newton method to use with implicit filtering is an area of active research
[56], [55]. Both SR1 and BFGS have been used, with SR1 performing modestly better in some
applications with bound constraints [270], [251], [271], [250], [55]. The implementation of
implicit filtering in the collection of MATLAB codes imfil.m uses BFGS as the default but
has SR1 as an option. We found BFGS with central differences to be consistently better in the
preparation of the (unconstrained!) computational examples in this book.
7.4 Implicit Filtering for Bound Constrained Problems

Implicit filtering was initially designed as an algorithm for bound constrained problems [250],
[120]. The bound constrained version we present here is simply a projected quasi-Newton
algorithm like the one presented in §5.5.3. There are other approaches to the implementation
and no best approach has emerged. We refer the reader to [120] and [55] for discussions of the
options.
We begin with scaling and the difference gradient. Central differences perform better, but
we do not evaluate f outside of the feasible region. Hence, if a point on the centered difference
stencil is outside of the feasible region, we use a one-sided difference in that direction. In order
to guarantee that at least one point in each direction is feasible, we scale the variables so that
Li = 0, Ui = 1, and h0 ≤ 1/2.
The sufficient decrease condition is (compare with (5.31))
(7.6) f (x(λ)) − f (x) ≤ α∇h f (x)T (x(λ) − x),
where
x(λ) = P(x − λ∇h f (x)).
One could terminate the iteration at a given scale when the analogue to (7.2)
(7.7)
x − x(1)
≤ τ h
holds or when
(7.8) f (xc ) < f (x ± rj ) for all x ± rj feasible,
which is the analogue to (6.7) for bound constrained problems.
Quasi-Newton methods for bound constraints can be constructed more simply for small
problems, like the ones to which implicit filtering is applied, where it is practical to store the
model of the inverse of the reduced Hessian as a full matrix. By using full matrix storage, the
complexity of bfgsrecb is avoided. One such alternative [53], [54], [55] to the updates in
§5.5.3 is to update the complete reduced Hessian and then correct it with information from the
new active set. This results in a two-stage update in which a model for the inverse of reduced
Hessian is updated with (4.5) to obtain

−1 sy T −1 ysT ssT
(7.9) R1/2 = I − T Rc I− T + T .
y s y s y s
Then the new reduced Hessian is computed using the active set information at the new point
−1 −1
(7.10) R+ = PA+ + PI+ R1/2 PI+ .
It is easy to show that Theorems 5.5.4 and 5.5.5 hold for this form of the update.
A FORTRAN implementation [119] of implicit filtering for bound constrained problems is
in the software collection. In the original version of that implementation a projected SR1 update
was used and a Cholesky factorization of the matrix R+ was performed to verify positivity. The
model Hessian was reinitialized to the identity whenever the scale or the active set changed.

7.5 Restarting and Minima at All Scales

No algorithm in this part of the book is guaranteed to find even a local minimum, much less
a global one. One approach to improving the robustness of these algorithms is to restart the
iteration after one sweep through the scales. A point x that is not changed after a call to
Algorithm imfilter1 (or imfilter2 or the bound constrained form of either) is called a
minimum at all scales.
If f satisfies (6.1), fs has a unique critical point that is also a local minimum that satisfies the
standard assumptions (and hence is a global minimum for fs ), and certain (strong!) technical
assumptions on the decay of φ near the minimum hold, then [120] a minimum at all scales is near
that global minimum of fs . In the unconstrained case this statement follows from the termination
criteria ((7.2) and (6.7)) for implicit filtering, Lemma 6.2.3 (or 6.2.7) and, if central differences
are used, Theorem 6.2.9. The analysis in [120] of the bound constrained case is more technical.
In practice, restarts are expensive and need not be done for most problems. However, restarts
have been reported to make a difference in some cases [178]. It is also comforting to know that
one has a minimum at all scales, and the author of this book recommends testing potential optima
with restarts before one uses the results in practice but not at the state where one is tuning the
optimizer or doing preliminary evaluation of the results.
7.6 Examples
Many of these examples are from [56]. For all the examples we report results with and without a
quasi-Newton Hessian. We report results for both forward and central differences. In the figures
the solid line corresponds to the BFGS Hessian, the dashed-dotted line to the SR1 Hessian, and
the dashed line to H = I, the steepest descent form of implicit filtering.
Unlike the smooth problems considered earlier, where convergence of the gradient to zero
was supported by theory, convergence of the simplex gradient to zero is limited by the noise in
the objective. We illustrate performance by plotting both the objective function value and the
norm of the simplex gradient. From these examples it is clear that the the graphs of function
value against the count of function evaluations is a better indicator of the performance of the
optimizer.
In all cases we terminated the iteration when either fdquasi had been called for each scale
or a budget of function evaluations had been exhausted. Once the code completes an iteration
and the number of function evaluations is greater than or equal to the budget, the iteration is
terminated.
The examples include both smooth and nonsmooth problems, with and without noise. A
serious problem for some algorithms of this type is their failure on very easy problems. For
most of the algorithms covered in this part of the book, we will present examples that illustrate
performance on this collection of problems.

The three Weber’s function examples all have minimizers at points at which the objective is
nondifferentiable. For the computations we used an initial iterate of (10, −10)T , a budget of
200 function evaluations, and {10 × 2−n }8n=−2 as the sequence of scales.
In each of the examples the performance of the two quasi-Newton methods was virtually
identical and far better than that without a quasi-Newton model Hessian. Forward and central
differences for the first two problems (Figures 7.1 and 7.2) perform almost equally well, with
forward having a slight edge. In Figure 7.3, however, the forward difference version of implicit
filtering finds a local minimum different from the global minimum that is located by central

Central Differences Central Differences

1
10 50
simplex gradient norm
100
function value
150
0
10
200
250
1
10 300
0 50 100 150 200 250 0 50 100 150 200 250
function evaluations function evaluations
Forward Differences Forward Differences

1
10 50
0 100
10 function value
150
1
10
200
2
10
250
3
10 300
0 50 100 150 200 250 0 50 100 150 200 250
Figure 7.1: First Weber Example

1
10 70
60
0 50
function value
10
40
30
1
10
20
10
2
10 0
0 50 100 150 200 250 0 50 100 150 200 250

1
10 70
60
0 50
function value
10
40
30
1
10
20
10
2
10 0
0 50 100 150 200 250 0 50 100 150 200 250
Figure 7.2: Second Weber Example


1
10 70
60
0
10
function value
50
1
10 40
30
2
10
20
3
10 10
0 50 100 150 200 0 50 100 150 200

2
10 70
60
0
10
function value
50
2
10 40
30
4
10
20
6
10 10
0 50 100 150 200 250 0 50 100 150 200 250
Figure 7.3: Third Weber Example
differencing. This, of course, is consistent with the theory, which does not claim that implicit
filtering is a global optimizer.
7.6.2 Parameter ID
We consider the parameter ID example from §1.6.2 using the data from §2.6.1. Recall that in this
example we use as data the values of the exact solution for c = k = 1 at the points ti = i/100
for 1 ≤ i ≤ 100. The initial iterate was (5, 5)T ; the sequence of scales was {2−k }12 k=1 . Implicit
filtering, like the globally convergent algorithms in the first part of the book, is fairly insensitive
to the choice of initial iterate, as we will see when we revisit this example in §8.5.2.
We report on both low (rtol = atol = 10−3 , Figure 7.4) and high (rtol = atol = 10−6 ,
Figure 7.5) accuracy computations. Note that after 200 function evaluations the function re-
duction from the central difference BFGS form of implicit filtering flattens out in both plots at
roughly the expected level of O(tol) while the other methods have not. This effect, which is
not uncommon, is one reason for our preference for the BFGS central difference form of the
algorithm.

The performance of the central difference BFGS form of implicit filtering should be very good,
since (see exercises 7.7.1 and 7.7.2) the difference approximation of the gradient is exact. We
would expect that good performance to persist in the perturbed case. We illustrate this with results
on two problems, both given by (6.12). One is an unperturbed problem (aj = bj = cj = 0
for all j) where H is a diagonal matrix with (H)ii = 1/(2i) for 1 ≤ i ≤ N . The other is a
perturbed problem with
ξ0 = (sin(1), sin(2), . . . , sin(N ))T , ξ1 = 0, ξ2 = (1, . . . , 1)T ,
a1 = a2 = .01, a3 = 0, b1 = (1, . . . , 1)T , b2 = 0, and c1 = c2 = 10π.


4 2
10 10
2
10
0
function value
10
0
10
2
10
2
10
4 4
10 10
0 50 100 150 200 250 0 50 100 150 200 250

5 2
10 10
0 function value 0
10 10
5 2
10 10
10 4
10 10
0 50 100 150 200 250 0 50 100 150 200 250
Figure 7.4: Parameter ID, tol = 10−3

4 2
10 10
0
10
2
function value
10
2
10
0
10
4
10
2 6
10 10
0 50 100 150 200 250 0 50 100 150 200 250

5 2
10 10
0 0
function value
10 10
5 2
10 10
10 4
10 10
0 50 100 150 200 250 0 50 100 150 200 250
Figure 7.5: Parameter ID, tol = 10−6


10 20
10 10
0 0
function value
10 10
10 20
10 10
20 40
10 10
0 100 200 300 400 0 100 200 300 400

10 5
10 10
0 0
function value
10 10
10 5
10 10
20 10
10 10
0 200 400 600 0 200 400 600
Figure 7.6: Unperturbed Quadratic, N = 4

2 5
10 10
0
10
0
function value
10
2
10
5
10
4
10
6 10
10 10
0 1000 2000 3000 4000 0 1000 2000 3000 4000

5 5
10 10
0 0
function value
10 10
5 5
10 10
10 10
10 10
0 1000 2000 3000 4000 0 1000 2000 3000 4000


2 2
10 10
0 0
function value
10 10
2 2
10 10
4 4
10 10
0 1000 2000 3000 4000 0 1000 2000 3000 4000

2 2
10 10
0
10
function value 0
10
2
10
2
10
4
10
6 4
10 10
0 100 200 300 400 0 100 200 300 400
Figure 7.8: Perturbed Quadratic, N = 4

2 4
10 10
1 2
10 10
function value
0 0
10 10
1 2
10 10
2 4
10 10
0 1000 2000 3000 4000 0 1000 2000 3000 4000

2 3
10 10
2
10
1
function value
10
1
10
0
10
0
10
1 1
10 10
0 1000 2000 3000 4000 0 1000 2000 3000 4000

If a3 = 0 then f may not return the same value when called with the same argument twice.
The reader is invited to explore the consequences of this in exercise 7.7.3.
The performance of the algorithms in this part of the book sometimes depends on the size of
the problem much more strongly than the Newton-based methods in Part I. In the case of implicit
filtering, that dependence is mostly a result of the cost of evaluation of the simplex gradient. To
illustrate this we consider our quadratic problems for N = 4 (Figures 7.6 and 7.8) and N = 32
(Figures 7.7 and 7.9).
For all the quadratic examples the initial iterate was
(1, 2, . . . , N )T
x0 =
10N
and the sequence of scales was {2−k }10
k=0 .
7.7 Exercises on Implicit Filtering

7.7.1. Let S be a nonsingular simplex. Show that DC (f : S) = f (x1 ) if f is a quadratic
function.
7.7.2. How would you expect forward and centered difference implicit filtering to perform when
applied to f (x) = xT x? Would the performance be independent of dimension? Test your
expectation with numerical experimentation.
7.7.3. Use implicit filtering to minimize the perturbed quadratic function with nonzero values of
a3 .
7.7.4. Try to solve the Lennard–Jones problem with implicit filtering for various values of M
and various initial iterates. Compare your best results with those in [142], [40], and [210].
Are you doing any better than you did in exercise 6.4.3?
7.7.5. Show that Theorems 5.5.4 and 5.5.5 hold if the projected BFGS update is implemented us-
ing (7.9) and (7.10). How would these formulas affect an implementation like bfgsrecb,
which is designed for problems in which full matrices cannot be stored?


Chapter 8
Direct Search Algorithms
In this chapter we discuss the class of direct search algorithms. These methods use values of
f taken from a set of sample points and use that information to continue the sampling. Unlike
implicit filtering, these methods do not explicitly use approximate gradient information. We will
focus on three such methods: the Nelder–Mead simplex algorithm [204], the multidirectional
search method [85], [261], [262], and the Hooke–Jeeves algorithm [145]. Each of these can be
analyzed using the simplex gradient techniques from Chapter 6. We will not discuss the very
general results based on the taxonomies of direct search methods from [263], [174], and [179] or
the recent research on the application of these methods to bound [173] or linear [175] constraints.
We include at the end of this chapter a short discussion of methods based on surrogate models
and a brief account of a very different search method, the DIRECT algorithm [150]. These two
final topics do not lead to algorithms that are easy to implement, and our discussions will be
very general with pointers to the literature.
8.1 The Nelder–Mead Algorithm

8.1.1 Description and Implementation
The Nelder–Mead [204] simplex algorithm maintains a simplex S of approximations to an
optimal point. In this algorithm the vertices {xj }N +1
j=1 are sorted according to the objective
function values
(8.1) f (x1 ) ≤ f (x2 ) ≤ · · · ≤ f (xN +1 ).
x1 is called the best vertex and xN +1 the worst. If several vertices have the same objective
value as x1 , the best vertex is not uniquely defined, but this ambiguity has little effect on the
performance of the algorithm.
The algorithm attempts to replace the worst vertex xN +1 with a new point of the form
(8.2) x(µ) = (1 + µ)x − µxN +1 ,
where x is the centroid of the convex hull of {xi }N

i=1
1 N
(8.3) x= xi .
N i=1
The value of µ is selected from a sequence
−1 < µic < 0 < µoc < µr < µe
135
by rules that we formally describe in Algorithm nelder. Our formulation of the algorithm
allows for termination if either f (xN +1 ) − f (x1 ) is sufficiently small or a user-specified number
of function evaluations has been expended.
Algorithm 8.1.1. nelder(S, f, τ, kmax)
1. Evaluate f at the vertices of S and sort the vertices of S so that (8.1) holds.
2. Set f count = N + 1.
3. While f (xN +1 ) − f (x1 ) > τ
(a) Compute x, (8.3), x(µr ), (8.2), and fr = f (x(µr )). f count = f count + 1.
(b) Reflect: If f count = kmax then exit. If f (x1 ) ≤ fr < f (xN ), replace xN +1 with
x(µr ) and go to step 3g.
(c) Expand: If f count = kmax then exit. If fr < f (x1 ) then compute fe = f (x(µe )).
f count = f count + 1. If fe < fr , replace xN +1 with x(µe ); otherwise replace
xN +1 with x(µr ). Go to to step 3g.
(d) Outside Contraction: If f count = kmax then exit. If f (xN ) ≤ fr < f (xN +1 ),
compute fc = f (x(µoc )). f count = f count + 1. If fc ≤ fr replace xN +1 with
x(µoc ) and go to step 3g; otherwise go to step 3f.
(e) Inside Contraction: If f count = kmax then exit. If fr ≥ f (xN +1 ) compute
fc = f (x(µic )). f count = f count + 1. If fc < f (xN +1 ), replace xN +1 with
x(µic ) and go to step 3g; otherwise go to step 3f.
(f) Shrink: If f count ≥ kmax − N , exit. For 2 ≤ i ≤ N + 1: set xi = x1 − (xi −
x1 )/2; compute f (xi ).
(g) Sort: Sort the vertices of S so that (8.1) holds.
A typical sequence [169] of candidate values for µ is
{µr , µe , µoc , µic } = {1, 2, 1/2, −1/2}.
Figure 8.1 is an illustration of the options in two dimensions. The vertices labeled 1, 2, and
3 are those of the original simplex.
The Nelder–Mead algorithm is not guaranteed to converge, even for smooth problems [89],
[188]. The failure mode is stagnation at a nonoptimal point. In §8.1.3 we will present some
examples from [188] that illustrate this failure. However, the performance of the Nelder–Mead
algorithm in practice is generally good [169], [274]. The shrink step is rare in practice and we
will assume in the analysis in §8.1.2 that shrinks do not occur. In that case, while a Nelder–Mead
iterate may not result in a reduction in the best function value, the average value
1 N +1
f= f (xj )
N +1 j=1
will be reduced.

SEARCH ALGORITHMS 137
1
r
oc
ic
3 2
Figure 8.1: Nelder–Mead Simplex and New Points
8.1.2 Sufficient Decrease and the Simplex Gradient

Our study of the Nelder–Mead algorithm is based on the simple ideas in §3.1. We will denote
the vertices of the simplex S k at the kth iteration by {xkj }N +1
j=1 . We will simplify notation by
k
suppressing explicit mention of S in what follows by denoting
V k = V (S k ), δ k = δ(f : S k ), K k = K(S k ), and Dk (f ) = D(f : S k ).
If V 0 is nonsingular then V k is nonsingular for all k > 0 [169]. Hence if S 0 is nonsingular so

is S k for all k and hence Dk (f ) is defined for all k.
We formalize this by assuming that our sequence of simplices satisfies the following assump-
tion.
Assumption 8.1.1. For all k,
• S k is nonsingular.
• The vertices satisfy (8.1).
• f k+1 < f k .
Assumption 8.1.1 is satisfied by the Nelder–Mead sequence if no shrink steps are taken
and the initial simplex directions are linearly independent [169]. The Nelder–Mead algorithm
demands that the average function value improve, but no control is possible on which value is
improved, and the simplex condition number can become unbounded.

We can define a sufficient decrease condition for search algorithms that is analogous to the
sufficient decrease condition for steepest descent and related algorithms (3.2). We will ask that
the k + 1st iteration satisfy
(8.4) f k+1 − f k < −α
Dk f
2 .
Here α > 0 is a small parameter. Our choice of sufficient decrease condition is motivated by the
smooth case and steepest descent, where (3.2) and the lower bound −λ̄ on λ from Lemma 3.2.3
lead to
f (xk+1 ) − f (xk ) ≤ −λ̄α
∇f (xk )
2 ,
which is a smooth form of (8.4). Unlike the smooth case, however, we have no descent direction
and must incorporate λ̄ into α. This leads to the possibility that if the simplex diameter is much
smaller than
Dk f
, (8.4) could fail on the first iterate. We address this problem with the scaling
σ+ (S 0 )
α = α0 .

D0 f
A typical choice in line search methods, which we use in our numerical results, is α0 = 10−4 .
The convergence result for smooth functions follows easily from Lemma 6.2.1.
Theorem 8.1.1. Let a sequence of simplices satisfy Assumption 8.1.1 and let the assumptions
of Lemma 6.2.1 hold, with the Lipschitz constants K k uniformly bounded. Assume that {f k } is
bounded from below. Then if (8.4) holds for all but finitely many k and
lim σ+ (S k )κ(V k ) = 0,
k→∞
then any accumulation point of the simplices is a critical point of f .

Proof. The boundedness from below of {f k } and (8.4) imply that f k → 0. Assumption 8.1.1
and (8.4) imply that limk→∞ Dk f = 0. Hence (6.2) implies

lim
∇f (xk1 )
≤ lim Kκ(V k )σ+ (S k ) +
Dk f
= 0.
k→∞ k→∞
∗
Hence, if x is any accumulation point of the sequence {xk1 } then ∇f (x∗ ) = 0. This completes
the proof since κ(V k ) ≥ 1 and therefore σ+ (V k ) → 0.
The result for the noisy functions that satisfy (6.1) with fs smooth reflects the fact that
the resolution is limited by the size of φ. In fact, if σ+ (S k ) is much smaller than
φ
S k , no
information on fs can be obtained by evaluating f at the vertices of S k and once σ+ (S k ) is
1/2
smaller than
φ
S k no conclusions on ∇fs can be drawn. If, however, the noise decays to zero
sufficiently rapidly near the optimal point, the conclusions of Theorem 8.1.1 still hold.
Theorem 8.1.2. Let a sequence of simplices satisfy Assumption 8.1.1 and let the assumptions
of Lemma 6.2.2 hold with the Lipschitz constants Ksk uniformly bounded. Assume that {f k } is
bounded from below. Then if (8.4) holds for all but finitely many k and if

φ
S k
(8.5) lim κ(V k ) σ+ (S k ) + = 0,
k→∞ σ+ (S k )
then any accumulation point of the simplices is a critical point of fs .
Proof. Our assumptions, as in the proof of Theorem 8.1.1, imply that Dk f → 0. Recall that
Lemma 6.2.2 implies that

φ
S k
(8.6)
Dk fs
≤
Dk f
+ K k κ(V k ) σ+ (S k ) + ,
σ+ (S k )
and the sequence {K k } is bounded because {Ksk } is. Hence, by (8.5), Dk fs → 0 as k → ∞.

Function Differences Max Oriented Length

2 2
10 10
0
10
0 10
2
2 10
10
4
10
4
10 6
10
6
10 10
8
8 10
10 10
0 50 100 150 0 50 100 150
Simplex Gradient Norm Simplex Condition

4 20
10 10
15
3
10
10
10
10
2
10 5
10
1 0
10 10
0 50 100 150 0 50 100 150
Figure 8.2: Unmodified Nelder–Mead, (τ, θ, φ) = (1, 15, 10)
8.1.3 McKinnon’s Examples

In this set of three examples from [188], N = 2, and

 θφ|(x)1 |τ + (x)2 + (x)22 , (x)1 ≤ 0,
f (x) =

θ(x)τ1 + (x)2 + (x)22 , (x)1 > 0.
The examples in [188] consider the parameter sets


 (3, 6, 400),
(τ, θ, φ) = (2, 6, 60),

(1, 15, 10).
The initial simplex was

√
x1 = (1, 1)T , x2 = (λ+ , λ− )T , x3 = (0, 0)T , where λ± = (1 ± 33)/8.
With this data, the Nelder–Mead iteration will stagnate at the origin, which is not a critical point
for f . The stagnation mode is repeated inside contractions that leave the best point (which is not
a minimizer) unchanged.
We terminated the iteration when the difference between the best and worst function values
was < 10−8 .
We illustrate the behavior of the Nelder–Mead algorithm in Figures 8.2, 8.3, and 8.4. In all
the figures we plot, as functions of the iteration index, the difference between the best and worst
function values, σ+ , the maximum oriented length, the norm of the simplex gradient, and the l2
condition number of the matrix of simplex directions. In all three problems stagnation is evident
from the behavior of the simplex gradients. Note also how the simplex condition number is
growing rapidly.


2 1
10 10
0
10
0 10
1
2 10
10
2
10
4
10 3
10
6
10 10
4
8 5
10 10
0 20 40 60 0 20 40 60

5 10
10 10
4 8
10 10
3 6
10 10
2 4
10 10
1 2
10 10
0 0
10 10
0 20 40 60 0 20 40 60
Function Differences Simplex Diameter

2 1
10 10
0
10 0
10
2
10
1
10
4
10
2
6 10
10
8 3
10 10
0 10 20 30 40 0 10 20 30 40

1 8
10 10
6
10
0 4
10 10
2
10
1 0
10 10
0 10 20 30 40 0 10 20 30 40

8.1.4 Restarting the Nelder–Mead Algorithm

When the Nelder–Mead iteration stagnates, a restart with the same best point and a different set
of directions can help sometimes. In order to formulate a restart scheme, one must first develop
a strategy for detecting stagnation. One might think that a large simplex condition would suffice
for this. However [204], the ability of the Nelder–Mead simplices to drastically vary their shape
is an important feature of the algorithm and looking at the simplex condition alone would lead
to poor results. Failure of (8.4), however, seems to indicate that something is wrong, and we
will use that as our stagnation detector.
Having detected stagnation, one must modify the simplex. Simply performing a shrink
step is not effective. The method we advocate here, from [155], is the oriented restart. The
motivation is that if the simplex gradient can be trusted to be in the correct orthant in RN , a new,
smaller simplex with orthogonal edges oriented with that quadrant should direct the iteration in
a productive direction.
We propose performing an oriented restart when (8.4) fails but f k+1 − f k < 0. This means
replacing the current simplex with vertices {xj }N +1
j=1 , ordered so that (8.1) holds, with a new
smaller simplex having vertices (before ordering!) {yj }N +1
j=1 with y1 = x1 and
(8.7) yj = y1 − βj−1 ej−1 for 2 ≤ j ≤ N + 1,
where, for 1 ≤ l ≤ N , el is the lth coordinate vector,


σ (S k )sign((Dk f )l ), (Dk f )l =
0,
1 −
βl =
2
σ− (S k ), (Dk f )l = 0,
and (Dk f )l is the lth component of Dk f . If Dk f = 0 we assume that the Nelder–Mead iteration
would have been terminated at iteration k because there is no difference between best and worst
values.
So, before ordering, the new simplex has the same first point as the old. The diameter of the
new simplex has not been increased since the diameter of the new simplex is at most σ+ (S k ).
Moreover all edge lengths have been reduced. So after reordering σ+ (S k+1 ) ≤ σ− (S k ). As for
κ, after the oriented shrink, but before reordering, κ(V ) = 1. After reordering, of course, the
best point may no longer be x1 . In any case the worst-case bound on κ is
√
(8.8) κ(V k+1 ) =
V k+1
2 ≤ (1 + N )2 .
In any case, the new simplex is well conditioned.

Returning to the McKinnon examples, we find that an oriented restart did remedy stagnation
for the smooth examples. The graphs in Figures 8.5, 8.6, and 8.7 report the same data as for the
unmodified algorithm, with stars on the plots denoting oriented restarts.
For the smoothest example, (τ, θ, φ) = (3, 6, 400), the modified form of Nelder–Mead took a
single oriented restart at the 21st iteration. For the less smooth of these two, (τ, θ, φ) = (2, 6, 60),
a single restart was taken on the 19th iteration. As one can see from Figures 8.6 and 8.7 the
restart had an immediate effect on the simplex gradient norm and overcame the stagnation.
For the nonsmooth example, (τ, θ, φ) = (1, 15, 10), in Figure 8.5, the modified algorithm
terminated with failure after restarting on the 44th, 45th, and 46th iterations. Since the objective
is not smooth at the stagnation point, this is the best we can expect and is far better than the
behavior of the unmodified algorithm, which stagnates with no warning of the failure.


2 2
10 10
0 0
10 10
2 2
10 10
4 4
10 10
0 10 20 30 40 50 0 10 20 30 40 50

3 8
10 10
6
10
2 4
10 10
2
10
1 0
10 10
0 10 20 30 40 50 0 10 20 30 40 50
Figure 8.5: Modified Nelder–Mead, (τ, θ, φ) = (1, 15, 10)

5 2
10 10
0 0
10 10
5 2
10 10
10 4
10 10
0 20 40 60 0 20 40 60

2 3
10 10
0 2
10 10
2 1
10 10
4 0
10 10
0 20 40 60 0 20 40 60


5 2
10 10
0 0
10 10
5 2
10 10
10 4
10 10
0 20 40 60 0 20 40 60

2 4
10 10
0 3
10 10
2 2
10 10
4 1
10 10
6 0
10 10
0 20 40 60 0 20 40 60
8.2 Multidirectional Search

One way to address the possible ill-conditioning in the Nelder–Mead algorithm is to require that
the condition numbers of the simplices be bounded. The multidirectional search (MDS) method
[85], [261], [262] does this by making each new simplex congruent to the previous one. The
results in this section, mostly taken from [29], show that MDS has convergence properties like
those of implicit filtering.
In the special case of equilateral simplices, V k is a constant multiple of V 0 and the simplex
condition number is constant. If the simplices are not equilateral, then κ(V ) may vary depending
on which vertex is called x1 , but (6.6) will hold in any case.
Figure 8.8 illustrates the two-dimensional case for two types of simplices. Beginning with
the ordered simplex S c with vertices x1 , x2 , x3 one first attempts a reflection step, leading to a
simplex S r with vertices x1 , r2 , r3 .
If the best function value of the vertices of S r is better than the best f (x1 ) in S 0 , S r is
(provisionally) accepted and expansion is attempted. The expansion step differs from that in
the Nelder–Mead algorithm because N new points are needed to make the new, larger simplex
similar to the old one. The expansion simplex S e has vertices x1 , e2 , e3 and is accepted over S r
if the best function value of the vertices of S e is better than the best in S r . If the best function
value of the vertices of S r is not better than the best in S c , then the simplex is contracted and
the new simplex has vertices x1 , c2 , c3 . After the new simplex is identified, the vertices are
reordered to create the new ordered simplex S + .
Similar to the Nelder–Mead algorithm, there are expansion and contraction parameters µe
and µc . Typical values for these are 2 and 1/2.
Algorithm 8.2.1. mds(S, f, τ, kmax)

Right Simplex
Equilateral Simplex
x3
x2
c3
x3
c2
c3 e2 r2
x1
x1 c2 x2
r3
r2
r3
e3
e3
e2
Figure 8.8: MDS Simplices and New Points
1. Evaluate f at the vertices of S and sort the vertices of S so that (8.1) holds.
2. Set f count = N + 1.
3. While f (xN +1 ) − f (x1 ) > τ
(a) Reflect: If f count = kmax then exit.

For j = 2, . . . , N +1: rj = x1 −(xj −x1 ); Compute f (rj ); f count = f count+1.
If f (x1 ) > minj {f (rj )} then goto step 3b else goto step 3c.
(b) Expand:
i. For j = 2, . . . , N + 1: ej = x1 − µe (xj − x1 ); Compute f (ej ); f count =
f count + 1.
ii. If minj {f (rj )} > minj {f (ej )} then
for j = 2, . . . N + 1: xj = ej
else
for j = 2, . . . N + 1: xj = rj
iii. Goto step 3d
(c) Contract: For j = 2, . . . , N + 1: xj = x1 + µc (xj − x1 ), Compute f (xj )
(d) Sort: Sort the vertices of S so that (8.1) holds.
If the function values at the vertices of S c are known, then the cost of computing S + is 2N
additional evaluations. Just as with the Nelder–Mead algorithm, the expansion step is optional
but has been observed to improve performance.
The extension of MDS to bound constrained and linearly constrained problems is not trivial.
We refer the reader to [173] and [175] for details.
8.2.2 Convergence and the Simplex Gradient

Assume that the simplices are either equilateral or right simplices (having one vertex from which
all N edges are at right angles). In those cases, as pointed out in [262], the possible vertices
created by expansion and reflection steps form a regular lattice of points. If the MDS simplices
remain bounded, only finitely many reflections and expansions are possible before every point
on that lattice has been visited and a contraction to a new maximal simplex size must take place.

This exhaustion of a lattice takes place under more general conditions [262] but is most clear for
the equilateral and right simplex cases.
Theorem 6.2.9 implies that infinitely many contractions and convergence of the simplex
diameters to zero imply convergence of the simplex gradient to zero. The similarity of The-
orem 6.2.9 to Lemma 6.2.2 and of Theorem 8.2.1, the convergence result for multidirectional
search, to Theorem 8.1.2 is no accident. The Nelder–Mead iteration, which is more aggressive
than the multidirectional search iteration, requires far stronger assumptions (well conditioning
and sufficient decrease) for convergence, but the ideas are the same. Theorems 6.2.9 and 8.2.1
can be used to extend the results in [262] to the noisy case. The observation in [85] that one
can apply any heuristic or machine-dependent idea to improve performance, say, by exploring
far away points on spare processors (the “speculative function evaluations” of [46]) without
affecting the analysis is still valid here.
Theorem 8.2.1. Let f satisfy (6.1) and assume that the set
{x | f (x) ≤ f (x01 )}
is bounded. Assume that the simplex shape is such that
(8.9) lim σ+ (S k ) → 0.
k→∞
Let B k be a ball of radius 2σ+ (S k ) about xk1 . Then if

φ
B k
lim =0
k→∞ σ+ (S k )
then every limit point of the vertices is a critical point of fs .
Recall that if the simplices are equilateral or right simplices, then (8.9) holds (see exer-
cise 8.6.2).
8.3 The Hooke–Jeeves Algorithm

The Hooke–Jeeves algorithm is like implicit filtering in that the objective is evaluated on a stencil
and the function values are used to compute a search direction. However, unlike implicit filtering,
there are only finitely many possible search directions and only qualitative information about
the function values is used.
The algorithm begins with a base point x and pattern size h, which is like the scale in implicit
filtering. In the next phase of the algorithm, called the exploratory move in [145], the function is
sampled at successive perturbations of the base point in the search directions {vj }, where vj is
the jth column of a direction matrix V . In [145] and our MATLAB implementation V = I. The
current best value fcb = f (xcb ) and best point xcb are recorded and returned. xcb is initialized
to x. The sampling is managed by first evaluating f at xcb + vj and only testing xcb − vj
if f (xcb + vj ) ≥ f (xcb ). The exploratory phase will either produce a new base point or fail
(meaning that xcb = x). Note that this phase depends on the ordering of the coordinates of x.
Applying a permutation to x could change the output of the exploration.
If the exploratory phase has succeeded, the search direction is
(8.10) dHJ = xcb − x
and the new base point is xcb . The subtle part of the algorithm begins here. Rather than center
the next exploration at xcb , which would use some of the same points that were examined in

the previous exploration, the Hooke–Jeeves pattern move step is aggressive and tries to move
further. The algorithm centers the next exploratory move at
xC = x + 2dHJ = xcb + dHJ .
If this second exploratory move fails to improve upon f (xcb ), then an exploratory move with
xcb as the center is tried. If that fails h is reduced, x is set to xcb , and the process is started over.
Note that when h has just been set, the base point and the center of the stencil for the exploratory
moves are the same, but afterward they are not.
If, after the first exploratory move, xcb = x (i.e., as it will be if x is the best point in the
pattern), then x is left unchanged and h is reduced.
Therefore, whenever h is reduced, the stencil centered at x has x itself as the best point.
This is exactly the situation that led to a shrink in the MDS algorithm and, as you might expect,
will enable us to prove a convergence result like those in the previous sections. In [145] h was
simply multiplied by a constant factor. Our description in Algorithm hooke follows the model
of implicit filtering and uses a sequence of scales. Choice of perturbation directions could be
generalized to any simplex shape, not just the right simplices used in [145].
Figure 8.9 illustrates the idea for N = 2. The base point x lies at the center of the stencil. If
f (x+ + − −
1 ) < f (x), f (x2 ) < f (x), f (x1 ) ≥ f (x), and f (x2 ) ≥ f (x),
then the new base point xb will be located above and to the right of x. The next exploratory
move will be centered at xC , which is the center of the stencil in the upper right corner of the
figure.
The reader, especially one who plans to implement this method, must be mindful that points
may be sampled more than once. For example, in the figure, if the exploratory move centered
at xC fails, f will be evaluated for the second time at the four points in the stencil centered
at xb unless the algorithm is implemented to avoid this. The MDS method is also at risk of
sampling points more than once. The implementations of Hooke–Jeeves and MDS in our suite
of MATLAB codes keep the most recent 4N iterations in memory to guard against this. This
reevaluation is much less likely for the Nelder–Mead and implicit filtering methods. One should
also be aware that the Hooke–Jeeves algorithm, like Nelder–Mead, does not have the natural
parallelism that implicit filtering and MDS do.
One could implement a variant of the Hooke–Jeeves iteration by using xC = x + dHJ
instead of xC = x + 2dHJ and shrinking the size of the simplex on stencil failure. This is the
discrete form of the classical coordinate descent algorithm [180] and can also be analyzed by
the methods of this section (see [279] for a different view).
Our implementation follows the model of implicit filtering as well as the description in
[145]. We begin with the exploratory phase, which uses a base point xb , base function value
fb = f (xb ), and stencil center xC . Note that in the algorithm xb = xC for the first exploration
and xC = xb +dHJ thereafter. Algorithm hjexplore takes a base point and a scale and returns
a direction and the value at the trial point x + d. We let V = I be the matrix of coordinate
directions, but any nonsingular matrix of search directions could be used. The status flag sf is
used to signal failure and trigger a shrink step.
Algorithm 8.3.1. hjexplore(xb , xC , f, h, sf )
1. fb = f (xb ); d = 0; sf = 0; xcb = xb ; fcb = f (xb ); xt = xC
2. for j = 1, . . . , N : p = xt + hvj ; if f (p) ≥ fb then p = xt − hvj ;

if f (p) < fb then xt = xcb = p; fb = f (xcb )
3. if xcb = xb ; sf = 1; xb = xcb

xC2+
xC1- xC1+
xC
x2+ xb xC2-
s1+
x1- x x1+
s1- s1+
s2-
x2-
Figure 8.9: Hooke–Jeeves Pattern and New Points
The exploration is coupled to the pattern move to complete the algorithm for a single value
of the scale. The inputs for Algorithm hjsearch are an initial iterate x, the function, and the
scale. On output, a point x is returned for which the exploration has failed. There are other
considerations, such as the budget for function evaluations, that should trigger a return from the
exploratory phase in a good implementation. In our MATLAB code hooke.m we pay attention
to the number of function evaluations and change in the function value as part of the decision to
return from the exploratory phase.
Algorithm 8.3.2. hjsearch(x, f, h)
1. xb = x; xC = x; sf = 1
2. Call hjexplore(x, xC , f, h, sf )
3. While sf = 1
(a) d = x − xb ; xb = x; xC = x + d
(b) Call hjexplore(x, xC , f, h, sf );
If sf = 0; xC = x; Call hjexplore(x, xC , f, h, sf )

Step 3b requires care in implementation. If sf = 0 on exit from the first call to hjexplore,
one should only test f at those points on the stencil centered at x that have not been evaluated
before.
The Hooke–Jeeves algorithm simply calls hjsearch repeatedly as h varies over a sequence
{hk } of scales.
Algorithm 8.3.3. hooke(x, f, {hk })
1. For k = 1, . . .
Call hjsearch(x, f, hk )
As is the case with implicit filtering, the Hooke–Jeeves algorithm can be applied to bound
constrained problems in a completely natural way [145], [227] by simply restricting the stencil
points to those that satisfy the bounds and avoiding pattern moves that leave the feasible region.
The Hooke–Jeeves algorithm shares with implicit filtering the property that extension to
bound constrained problems is trivial [145]. One simply restricts the exploratory and pattern
moves to the feasible set.
8.3.2 Convergence and the Simplex Gradient

As with MDS, if the set of sampling points remains bounded, only finitely many explorations
can take place before hjsearch returns and the scale must be reduced. The conditions for
reduction in the scale include failure of an exploratory move centered at the current best point x.
This means that we can apply Theorem 6.2.9 with κ+ = 1 to prove the same result we obtained
for MDS.
Theorem 8.3.1. Let f satisfy (6.1). Let {xk } be the sequence of Hooke–Jeeves best points.
Assume that the set
{x | f (x) ≤ f (x0 )}
is bounded. Then let hk → 0 and if

φ
B k
lim = 0,
k→∞ σ+ (S k )
where B k is the ball of radius 2hk about xk , then every limit point of {xk } is a critical point of
fs .
8.4 Other Approaches

In this section we briefly discuss two methods that have been used successfully for noisy prob-
lems. These methods are substantially more difficult to implement than the ones that we have
discussed so far and we will give few details. The pointers to the literature are a good starting
place for the interested and energetic reader.
8.4.1 Surrogate Models

As any sampling method progresses, the function values can be used to build a (possibly)
quadratic model based, for example, on interpolation or least squares fit-to-data. Such mod-
els are called surrogates or response surfaces. Even for smooth f there are risks in doing this.
Points from early in the iteration may corrupt an accurate model that could be built from the
more recent points; however, the most recent points alone may not provide a rich enough set of

interpolatory data. The function being modeled could be too complex to be modeled in a sim-
ple way (think of the Lennard–Jones function), and very misleading results could be obtained.
However, this approach is often very productive even for smooth problems in which evaluation
of f is very expensive (see [28] for a high-flying example).
Initialization of the model requires an initial set of points at which to sample f . Selection of
this point set is not a trivial issue, and the regular stencils used in implicit filtering and the direct
search algorithms are very poor choices. The study of this issue alone is a field in itself, called
design and analysis of computer experiments (DACE) [27], [167], [230].
Having built such a model, one then finds one or more local minima of the model. One can
use either a conventional gradient-based method, a sampling algorithm of the type discussed in
Chapters 7 or 8, or an algorithm that is itself based on building models like the one described in
[62], the nongradient-based approaches being used when the model is expected to be multimodal
or nonconvex. Upon minimizing the model, one then evaluates f again at one or more new points.
The implementation of such a scheme requires careful coordination between the sampling
of the function, the optimization of the model, and the changing of the set of sample points. We
refer the reader to [28] and [4] for more information on recent progress in this area.
8.4.2 The DIRECT Algorithm

Suppose f is a Lipschitz continuous function on [a, b] with Lipschitz constant L. If one has a
priori knowledge of L, one can use this in a direct search algorithm to eliminate intervals of
possible optimal points based on the function values at the endpoints of these intervals. The
Shubert algorithm [146], [214], [241] is the simplest way to use this idea. The method begins
with the fact that
(8.11) f (x) ≥ flow (x, a, b) = max(f (a) − L(x − a), f (b) − L(b − x))
for all x ∈ [a, b]. If one samples f repeatedly, one can use (8.11) on a succession of intervals
and obtain a piecewise linear approximation to f . If In = [an , bn ] ⊂ [a, b] then f (x) ≥
flow (x, an , bn ) on In , the minimum value of flow (x, an , bn ) is
Vn = (f (an ) + f (bn ) − L(bn − an ))/2,
and the minimizer is
Mn = (f (an ) − f (bn ) + L(bn + an ))/(2L).
The algorithm begins with I0 = [a, b], selects the interval for which Vn is least, and divides at
Mn . This means that if K intervals have been stored we have, replacing In and adding IK+1 to
the list,
In = [an , Mn ] and IK+1 = [Mn , bn ].
The sequence of intervals is only ordered by the iteration counter, not by location. In this way
the data structure for the intervals is easy to manage.
If there are p and k such that p = k and Vp ≥ max(f (ak ), f (bk )), then Ip need not be
searched any longer, since the best value from Ip is worse than the best value in Ik . The
algorithm’s rule for division automatically incorporates this information and will never sample
from Ip .
There are two problems with this algorithm. One cannot expect to know the Lipschitz
constant L, so it must be estimated. An estimated Lipschitz constant that is too low can lead to
erroneous rejection of an interval. An estimate that is too large will lead to slow convergence,
since intervals that should have been discarded will be repeatedly divided. The second problem

10
0
0 1 2 3 4 5 6 7
Figure 8.10: Selection of Intervals in DIRECT
is far more serious. The obvious generalization of the Shubert algorithm to more than one
dimension would replace intervals by N -dimensional hyperrectangles and require sampling at
each of the 2N vertices of the rectangle to be divided. This exponential complexity makes this
trivial generalization of the Shubert algorithm completely impractical.
The DIRECT algorithm [150] attempts to address these problems by sampling at the midpoint
of the hyperrectangle rather than the vertices and indirectly estimating the Lipschitz constant as
the optimization progresses. The scheme is not completely successful in that the mesh of sample
points becomes everywhere dense as the optimization progresses. Hence the algorithm becomes
an exhaustive search, a fact that is used in [150] to assert global convergence. In spite of the
exponential complexity of exhaustive search, even one with a fixed-size mesh (a problem with
any deterministic algorithm that is truly global [248]), DIRECT has been reported to perform well
in the early phases of the iteration [150], [108] and for suites of small test problems. DIRECT is
worth consideration as an intermediate algorithmic level between methods like implicit filtering,
Nelder–Mead, Hooke–Jeeves, or MDS on the conservative side and nondeterministic methods
like simulated annealing or genetic algorithms on the radical side.
We will describe DIRECT completely only for the case N = 1. This will make clear how
the algorithm implicitly estimates the Lipschitz constant. The extension to larger values of N
requires careful management of the history of subdivision of the hyperrectangles, and we will give
a simple pictorial account of that. For more details we refer to [150], [147], or the documentation
[108] of the FORTRAN implementation of DIRECT from the software collection.
As with the Shubert algorithm we begin with an interval [a, b] but base our lower bound and
our subdivision strategy on the midpoint c = (a + b)/2. If the Lipschitz constant L is known

then
f (x) ≥ f (c) − L(b − a)/2.
If we are to divide an interval and also retain the current value c as the midpoint of an interval
in the set of intervals, we must divide an interval into three parts. If there are K intervals on
the list and an interval In = [an , bn ] with midpoint cn has been selected for division, the new
intervals are
IK+1 = [an , an + (bn − an )/3], In = [an + (bn − an )/3, bn − (bn − an )/3], and
IK+2 = [bn − (bn − an )/3, bn ].
So cn is still the midpoint of In and two new midpoints have been added.
The remaining part of the algorithm is the estimation of the Lipschitz constant and the
simultaneous selection of the intervals to be divided. If the Lipschitz constant were known, an
interval would be selected for division if f (c) − L(b − a)/2 were smallest. This is similar to
the Shubert algorithm. In order for there to even exist a Lipschitz constant that would force
an interval to be selected for division in this way, that interval must have the smallest midpoint
value of all intervals having the same length. Moreover, there should be no interval of a different
length for which f (c) − L(b − a)/2 was smaller.
The DIRECT algorithm applies this rule to all possible combinations of possible Lipschitz
constants and interval sizes. If one plots the values of f at the midpoints against the lengths of
the intervals in the list to obtain a plot like the one in Figure 8.10, one can visually eliminate
all but one interval for each interval length. By taking the convex hull of the lowest points, one
can eliminate interval lengths for which all function values are so high that f (c) − L(b − a)/2
would be smaller for the best point at a different length no matter what L was. For example, the
three points that intersect the line in Figure 8.10 would correspond to intervals that would be
subdivided at this step. The slopes of the line segments through the three points are estimates
of the Lipschitz constant. These estimates are not used explicitly, as they would be in the
Shubert algorithm, but implicitly in the process of selection of intervals to be divided. Unlike
the Shubert algorithm, where the Lipschitz constant is assumed known, the DIRECT algorithm
will eventually subdivide every interval.
The resulting algorithm may divide more than a single interval at each stage and the number
of intervals to be divided may vary. This is easy to implement for a single variable. However,
for more than one variable there are several ways to divide a hyperrectangle into parts and one
must keep track of how an interval has previously been divided in order not to cluster sample
points prematurely by repeatedly dividing an interval in the same way. Figures 8.11 and 8.12,
taken from [108], illustrate this issue for N = 2. In Figure 8.11 the entire rectangle will be
divided. Shading indicates that the rectangle has been selected for division. Four new midpoints
are sampled. The subdivision into new rectangles could be done in two ways: the figure shows
an initial horizontal subdivision followed by a vertical division of the rectangle that contains the
original center. The second division is shown in Figure 8.12. The two shaded rectangles are
selected for division. Note that four new centers are added to the small square and two to the
larger, nonsquare, rectangle. In this way the minimum number of new centers is added.
DIRECT parallelizes in a natural way. All hyperrectangles that are candidates for division
may be divided simultaneously, and for each hyperrectangle the function evaluations at each of
the new midpoints can also be done in parallel. We refer the reader to [150] and [108] for details
on the data structures and to [108] for a FORTRAN implementation and additional discussion
on the exploitation of parallelism.

6 6 6
5 9 8 5 9 8 5 9 8
2 2 2
Figure 8.11: Initial Division of Rectangles with DIRECT
6 6 6
9 9
5 9 8 7 5 6 9 8 7 5 6 9 8
4 4
2 3 2 6 3 2 6
Figure 8.12: Second Division of Rectangles with DIRECT
8.5 Examples
In each of the examples we compare the central difference BFGS form of implicit filtering from
§7.6 (solid line) with the Nelder–Mead (dashed line), Hooke–Jeeves (solid line with circles), and
MDS (dashed-dotted line) algorithms.
For each example we specified both an initial iterate and choice of scales. This is sufficient
to initialize both implicit filtering and Hooke–Jeeves. We used the implicit filtering forward
difference stencil as the initial simplex for both Nelder–Mead and MDS.
The plots reflect the differences in the startup procedures for the varying algorithms. In
particular, Nelder–Mead and MDS sort the simplex and hence, if the initial iterate is not the best
point, report the lower value as the first iterate.
The relative performance of the various methods on these example problems should not be
taken as a definitive evaluation, nor should these examples be thought of as a complete suite of test
problems. One very significant factor that is not reflected in the results in this section is that both
implicit filtering [69], [55] and multidirectional search [85] are easy to implement in parallel,
while Nelder–Mead and Hooke–Jeeves are inherently sequential. The natural parallelism of
implicit filtering and multidirectional search can be further exploited by using idle processors to
explore other points on the line search direction or the pattern.

The initial data were the same as that in §7.6.1. Implicit filtering does relatively poorly for this
problem because of the nonsmoothness at optimality. The resluts for these problems are plotted

−80
−100
−120
−140
−160
function value
−180
−200
−220
−240
−260
−280
0 50 100 150 200 250
function evaluations
Figure 8.13: First Weber Example
in Figures 8.13, 8.14, and 8.15. The other three algorithms perform equally well. Note in the
third example that MDS finds a local minimum that is not the global minimum.
8.5.2 Parameter ID
In the computations reported in this section each algorithm was allowed 500 evaluations of f
and the sequence of scales was {2−j }12 j=1 .
We begin with the two examples from §7.6.2. With the initial iterate of (5, 5)T , the exact
solution to the continuous problem lies on the grid that the Hooke–Jeeves algorithm uses to search
for the solution. This explains the unusually good performance of the Hooke–Jeeves optimization
shown in both Figures 8.16 and 8.17. When the initial iterate is changed to (5.1, 5.3)T , the
performance of Hooke–Jeeves is very different as one can see from Figures 8.18 and 8.19. The
other algorithms do not have such a sensitivity to the initial iterate for this example. We have no
explanation for the good performance turned in by the Nelder–Mead algorithm on this problem.

The problems and initial data are the same as those in §7.6.3. This is an example of how sampling
algorithms can perform poorly for very simple problems and how this poor performance is made
worse by increasing the problem size. Exercise 7.7.4 illustrates this point very directly. One
would expect implicit filtering to do well since a central difference gradient has no error for
quadratic problems. For the larger problem (N = 32, Figures 8.21 and 8.23), both the Nelder–
Mead and MDS algorithms perform poorly while the Hooke–Jeeves algorithm does surprisingly
well. The difference in performance of the algorithms is much smaller for the low-dimensional
problem (N = 4, Figures 8.20 and 8.22).

70
60
50
function value
40
30
20
10
0
0 20 40 60 80 100 120 140
Figure 8.14: Second Weber Example
70
60
50
function value
40
30
20
10
0 20 40 60 80 100 120
Figure 8.15: Third Weber Example

2
10
1
10
0
10
function value
−1
10
−2
10
−3
10
0 100 200 300 400 500 600
Figure 8.16: Parameter ID, tol = 10−3 , x0 = (5, 5)T
2
10
0
10
−2
10
function value
−4
10
−6
10
−8
10
−10
10
0 100 200 300 400 500 600
Figure 8.17: Parameter ID, tol = 10−6 , x0 = (5, 5)T

2
10
1
10
0
10
function value
−1
10
−2
10
−3
10
0 100 200 300 400 500 600
Figure 8.18: Parameter ID, tol = 10−3 , x0 = (5.1, 5.3)T
2
10
1
10
0
10
−1
10
−2
10
function value
−3
10
−4
10
−5
10
−6
10
−7
10
−8
10
0 100 200 300 400 500 600
Figure 8.19: Parameter ID, tol = 10−6 , x0 = (5.1, 5.3)T

2
10
1
10
0
10
−1
10
−2
10
function value
−3
10
−4
10
−5
10
−6
10
−7
10
−8
10
0 50 100 150 200 250 300 350 400
4
10
2
10
0
10
function value
−2
10
−4
10
−6
10
−8
10
0 500 1000 1500 2000 2500 3000 3500

2
10
1
10
0
10
−1
function value
10
−2
10
−3
10
−4
10
−5
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
3
10
2
10
1
10
0
function value
10
−1
10
−2
10
−3
10
−4
10
0 500 1000 1500 2000 2500 3000 3500

8.6 Exercises on Search Algorithms

8.6.1. Let Sl for 1 ≤ l ≤ 3 be the simplices having one vertex at (xl )1 = (10, 10, 10, 10) and
direction vectors V l given by
V 1 = diag(1, 2, 3, 4), V2 = diag(4, 3, 2, 1), V3 = diag(2, 2, 2, 2).
For each l = 1, 2, 3, apply the Nelder–Mead algorithm to the function f defined for x ∈ R4
by
(x1 − x2 x3 x4 )2 + (x2 − x3 x4 )2 + (x3 − x4 )2 + x24
with the initial simplex V l . What happened? This example is one of Nelder’s favorites
[203].
8.6.2. Show that if the set {x | f (x) ≤ f (x01 )} is bounded and S0 is either an equilateral or a
right simplex, then (8.9) holds.
8.6.3. One can modify MDS [260] by eliminating the expansion step and only computing re-
flected points until one is found that is better than x1 . If no reflected points are better,
then perform a contraction step. Prove that Theorem 8.2.1 holds for this implementation.
Implement MDS in this way and compare it with Algorithm mds. Are the savings in calls
to f for each iterate realized in a savings for the entire optimization?
8.6.4. The easiest problem in optimization is to minimize xT x. Give the algorithms in this
section a chance to show what they can do by using them to solve this problem. Try several
initial iterates (or initial simplices/patterns) and several problem dimensions (especially
N = 8, 16, 32).
8.6.5. The search methods in this section impose a structure on the sampling and thereby hope
to find a useful optimal point far more efficiently than using an unstructured deterministic
or random search. Implement an unstructured search and use your algorithm to minimize
xT x when N = 2. For an example of such a method, take the one from [6], please.
8.6.6. The Spendley, Hext, and Himsworth algorithm [244] manages the simplices in a very
different way from those we’ve discussed in the text. Use the information in [244] and
[267] to implement this algorithm. Use Theorem 6.2.9 to prove convergence for N = 2.
What happens to both your implementation and analysis when you try N = 3 or arbitrary
N ? Explain Table 5 in [244].
8.6.7. Use any means necessary to solve the Lennard–Jones problem. Have your results improved
since you tried exercises 6.4.3 and 7.7.4?


Bibliography
[1] E. Aarts and J. Korst, Simulated Annealing and Boltzmann Machines, Wiley, New
York, 1989.
[2] L. Adams and J. L. Nazareth, eds., Linear and Nonlinear Conjugate Gradient-Related
Methods, SIAM, Philadelphia, 1996.
[3] M. Al-Baali, Descent property and global convergence of the Fletcher-Reeves method
with inexact line searches, IMA J. Numer. Anal., 5 (1985), pp. 121–124.
[4] N. Alexandrov, J. E. Dennis, R. M. Lewis, and V. Torczon, A Trust Region Frame-
work for Managing the Use of Approximation Models in Optimization, Tech. Rep. 97-50,
Institute for Computer Applications in Science and Engineering, Hampton, VA, October
1997.
[5] E. L. Allgower and K. Georg, Numerical path following, in Handbook of Numerical
Analysis, P. G. Ciarlet and J. L. Lions, eds., vol. 5, North–Holland, Amsterdam, 1997,
pp. 3–207.
[6] Anonymous, A new algorithm for optimization, Math. Programming, 3 (1972), pp. 124–
128.
[7] L. Armijo, Minimization of functions having Lipschitz-continuous first partial derivatives,
Pacific J. Math., 16 (1966), pp. 1–3.
[8] U. M. Ascher and L. R. Petzold, Computer Methods for Ordinary Differential Equa-
tions and Differential-Algebraic Equations, SIAM, Philadelphia, 1998.
[9] K. E. Atkinson, Iterative variants of the Nyström method for the numerical solution of
integral equations, Numer. Math., 22 (1973), pp. 17–31.
[10] B. M. Averick and J. J. Moré, User Guide for the MINPACK-2 Test Problem Collection,
Tech. Rep. ANL/MCS-TM-157, Math. and Comp. Science Div. Report, Argonne National
Laboratory, Argone, IL, October 1991.
[11] O. Axelsson, Iterative Solution Methods, Cambridge University Press, Cambridge, 1994.
[12] S. Banach, Sur les opérations dans les ensembles abstraits et leur applications aux
équations intégrales, Fund. Math., 3 (1922), pp. 133–181.
[13] H. T. Banks and H. T. Tran, Mathematical and Experimental Modeling of Physical
Processes, unpublished notes for MA 573-4, 1997, Department of Mathematics, North
Carolina State University, Raleigh, NC.
[14] M. S. Barlett, An inverse matrix adjustment arising in discriminant analysis, Ann. Math.
Stat., 22 (1951), pp. 107–111.
161
162 BIBLIOGRAPHY
[15] R. Barrett, M. Berry, T. F. Chan, J. Demmel, J. Donato, J. Dongarra, V. Eijkhout,

R. Pozo, C. Romine, and H. van der Vorst, Templates for the Solution of Linear
Systems: Building Blocks for Iterative Methods, SIAM, Philadelphia, 1994.
[16] K. J. Bathe and A. P. Cimento, Some practical procedures for the solution of nonlinear
finite element equations, Comput. Methods. Appl. Mech. Engrg., 22 (1980), pp. 59–85.
[17] A. Ben-Israel, A Newton-Raphson method for the solution of systems of equations, J.

Math. Anal. Appl., 15 (1966), pp. 243–252.
[18] D. P. Bertsekas, On the Goldstein-Levitin-Polyak gradient projection method, IEEE

Trans. Automat. Control, 21 (1976), pp. 174–184.
[19] , Projected Newton methods for optimization problems with simple constraints,
SIAM J. Control Optim., 20 (1982), pp. 221–246.
[20] J. T. Betts, An improved penalty function method for solving constrained parameter
optimization problems, J. Optim. Theory Appl., 16 (1975), pp. 1–24.
[21] , Solving the nonlinear least square problem: Application of a general method, J.
Optim. Theory Appl., 18 (1976), pp. 469–483.
[22] J. T. Betts, M. J. Carter, and W. P. Huffman, Software for Nonlinear Optimization,

Tech. Rep. MEA-LR-083 R1, Mathematics and Engineering Analysis Library Report,
Boeing Information and Support Services, Seattle, WA, June 6, 1997.
[23] J. T. Betts and P. D. Frank, A sparse nonlinear optimization algorithm, J. Optim. Theory
Appl., 82 (1994), pp. 519–541.
[24] P. T. Boggs, The convergence of the Ben-Israel iteration for nonlinear least squares
problems, Math. Comp., 30 (1976), pp. 512–522.
[25] P. T. Boggs and J. E. Dennis, A stability analysis for perturbed nonlinear iterative
methods, Math. Comp., 30 (1976), pp. 1–17.
[26] I. Bongatz, A. R. Conn, and P. L. Toint, CUTE: Constrained and unconstrained testing
environment, ACM Trans. Math. Software, 21 (1995), pp. 123–160.
[27] A. J. Booker, DOE for Computer Output, Tech. Rep. BCSTECH-94-052, Boeing Com-
puter Services, Seattle, WA, 1994.
[28] A. J. Booker, J. E. Dennis, P. D. Frank, D. B. Serafini, V. Torczon, and M. W.

Trosset, A rigorous framework for optimization of expensive function by surrogates,
Structural Optimization, 17 (1999), pp. 1–13.
[29] D. M. Bortz and C. T. Kelley, The simplex gradient and noisy optimization problems,
in Computational Methods in Optimal Design and Control, J.T. Borggaard, J. Burns,
E. Cliff, S. Schrenk, eds., Birkhauser, Boston, 1998, pp. 77–90.
[30] A. Bouaricha, Tensor methods for large, sparse, unconstrained optimization, SIAM J.
Optim., 7 (1997), pp. 732–756.
[31] H. Brakhage, Über die numerische Behandlung von Integralgleichungen nach der
Quadraturformelmethode, Numer. Math., 2 (1960), pp. 183–196.
[32] K. E. Brenan, S. L. Campbell, and L. R. Petzold, The Numerical Solution of Initial

Value Problems in Differential-Algebraic Equations, SIAM, Philadelphia, 1996.

BIBLIOGRAPHY 163
[33] P. N. Brown and Y. Saad, Convergence theory of nonlinear Newton–Krylov algorithms,

SIAM J. Optim., 4 (1994), pp. 297–330.
[34] C. G. Broyden, A class of methods for solving nonlinear simultaneous equations, Math.
Comp., 19 (1965), pp. 577–593.
[35] , Quasi-Newton methods and their application to function minimization, Math.

Comp., 21 (1967), pp. 368–381.
[36] , A new double-rank minimization algorithm, AMS Notices, 16 (1969), p. 670.
[37] C. G. Broyden, J. E. Dennis, and J. J. Moré, On the local and superlinear convergence
of quasi-Newton methods, J. Inst. Math. Appl., 12 (1973), pp. 223–246.
[38] R. H. Byrd, T. Derby, E. Eskow, K. P. B. Oldenkamp, and R. B. Schnabel, A new

stochastic/perturbation method for large-scale global optimization and its application to
water cluster problems, in Large Scale Optimization: State of the Art, W. W. Hager, D. W.
Hearn, and P. Pardalos, eds., Kluwer Academic Publishers B.V., Boston, 1994, pp. 68–81.
[39] R. H. Byrd, C. L. Dert, A. H. G. R. Kan, and R. B. Schnabel, Concurrent stochastic

methods for global optimization, Math. Programminng., 46 (1990), pp. 1–30.
[40] R. H. Byrd, E. Eskow, and R. B. Schnabel, A New Large-Scale Global Optimiza-

tion Method and Its Application to Lennard-Jones Problems, Tech. Rep. CU-CS-630-92,
University of Colorado at Boulder, November 1992.
[41] R. H. Byrd, H. F. Khalfan, and R. B. Schnabel, Analysis of a symmetric rank-one

trust region method, SIAM J. Optim., 6 (1996), pp. 1025–1039.
[42] R. H. Byrd, P. Lu, J. Nocedal, and C. Zhu, A limited memory algorithm for bound
constrained optimization, SIAM J. Sci. Comput., 16 (1995), pp. 1190–1208.
[43] R. H. Byrd and J. Nocedal, A tool for the analysis of quasi-Newton methods with
application to unconstrained minimization, SIAM J. Numer. Anal., 26 (1989), pp. 727–
739.
[44] 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.
[45] R. H. Byrd, J. Nocedal, and Y. Yuan, Global convergence of a class of quasi-Newton

methods on convex problems, SIAM J. Numer. Anal., 24 (1987), pp. 1171–1190.
[46] R. H. Byrd, R. B. Schnabel, and G. A. Schultz, Parallel quasi-Newton methods for

unconstrained optimization, Math. Programming, 42 (1988), pp. 273–306.
[47] P. H. Calamai and J. Moré, Projected gradient methods for linearly constrained prob-
lems, Math. Programming, 39 (1987), pp. 93–116.
[48] S. L. Campbell, C. T. Kelley, and K. D. Yeomans, Consistent initial conditions for

unstructured higher index DAEs: A computational study, in Proc. Conference on Compu-
tational Engineering in Systems Applications (CESA’96), Lille, France, 1996, pp. 416–
421.
[49] S. L. Campbell and C. D. Meyer, Generalized Inverses of Linear Transformations,

Dover Press, New York, 1991.

164 BIBLIOGRAPHY
[50] S. L. Campbell and K. D. Yeomans, Behavior of the nonunique terms in general DAE
integrators, Appl. Numer. Math., (1998), to appear.
[51] R. G. Carter, On the global convergence of trust region algorithms using inexact gradient
information, SIAM J. Numer. Anal., 28 (1991), pp. 251–265.
[52] A. Cauchy, Methode generale pour la resolution des systemes d’equations simultanees,
Comp. Rend. Acad. Sci. Paris, (1847), pp. 536–538.
[53] T. D. Choi. Private Communication, 1998.
[54] , Bound Constrained Optimization, PhD thesis, North Carolina State University,
Raleigh, North Carolina, 1999.
[55] T. D. Choi, O. J. Eslinger, C. T. Kelley, J. W. David, and M. Etheridge, Optimization

of automotive valve train components with implict filtering, Optim. Engrg., 1 (2000),
pp. 9–28.
[56] T. D. Choi and C. T. Kelley, Superlinear convergence and implicit filtering, SIAM J.
Optim., 10 (2000), pp. 1149–1162.
[57] T. F. Coleman and Y. Li, On the convergence of interior-reflective Newton methods for
nonlinear minimization subject to bounds, Math. Programming, 67 (1994), pp. 189–224.
[58] , An interior trust region approach for nonlinear minimization subject to bounds,
SIAM J. Optim., 6 (1996), pp. 418–445.
[59] T. F. Coleman and J. J. Moré, Estimation of sparse Jacobian matrices and graph
coloring problems, SIAM J. Numer. Anal., 20 (1983), pp. 187–209.
[60] 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 Com-
putations, J. R. Bunch and D. J. Rose, eds., Academic Press, NewYork, 1976, pp. 309–332.
[61] A. R. Conn, , K. Scheinberg, and P. L. Toint, On the convergence of derivative-free

methods for unconstrained optimization, in Approximation Theory and Optimization:
Tributes to M. J. D. Powell, A. Iserles and M. Buhmann, eds., Cambridge University
Press, 1997, pp. 83–108.
[62] , Recent progress in unconstrained optimization without derivatives, Math. Pro-

gramming Ser. B, 79 (1997), pp. 397–414.
[63] A. R. Conn, N. I. M. Gould, and P. L. Toint, Global convergence of a class of trust

region algorithms for optimization problems with simple bounds, SIAM J. Numer. Anal.,
25 (1988), pp. 433–460.
[64] , Testing a class of methods for solving minimization problems with simple bounds
on the variables, Math. Comp., 50 (1988), pp. 399–430.
[65] , Convergence of quasi-Newton matrices generated by the symmetric rank one up-
date, Math. Programming Ser. A, 50 (1991), pp. 177–195.
[66] , LANCELOT: A Fortran Package for Large-Scale Nonlinear Optimization (Release

A), Springer Series in Computational Mathematics, Springer-Verlag, Heidelberg, Berlin,
New York, 1992.

BIBLIOGRAPHY 165
[67] 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.
[68] J. W. Daniel, The conjugate gradient method for linear and nonlinear operator equations,
SIAM J. Numer. Anal., 4 (1967), pp. 10–26.
[69] J. W. David, C. Y. Cheng, T. D. Choi, C. T. Kelley, and J. Gablonsky, Optimal Design

of High Speed Mechanical Systems, Tech. Rep. CRSC-TR97-18, Center for Research in
Scientific Computation, North Carolina State University, Raleigh, July 1997; Math.
Modelling Sci. Comput., to appear.
[70] J. W. David, C. T. Kelley, and C. Y. Cheng, Use of an Implicit Filtering Algorithm

for Mechanical System Parameter Identification. SAE Paper 960358, 1996 SAE Interna-
tional Congress and Exposition Conference Proceedings, Modeling of CI and SI Engines,
Society of Automotive Engineers, pp. 189–194.
[71] W. C. Davidon, Variable Metric Methods for Minimization, Tech. Rep. ANL-5990, Ar-
gonne National Laboratory, Argone, IL, 1959.
[72] , Variable metric method for minimization, SIAM J. Optim., 1 (1991), pp. 1–17.
[73] T. J. Dekker, Finding a zero by means of successive linear interpolation, in Constructive

Aspects of the Fundamental Theorem of Algebra, B. Dejon and P. Henrici, eds., Wiley-
Interscience, New York, 1969, pp. 37–48.
[74] R. S. Dembo, S. C. Eisenstat, and T. Steihaug, Inexact Newton methods, SIAM J.

Numer. Anal., 19 (1982), pp. 400–408.
[75] R. Dembo and T. Steihaug, Truncated Newton algorithms for large-scale optimization,
Math. Programming, 26 (1983), pp. 190–212.
[76] J. E. Dennis, Nonlinear least squares and equations, in The State of the Art in Numerical
Analysis, D. Jacobs, ed., Academic Press, London, 1977, pp. 269–312.
[77] J. E. Dennis, D. M. Gay, and R. E.Welsch, An adaptive nonlinear least-squares algo-

rithm, ACM Trans. Math. Software, 7 (1981), pp. 348–368.
[78] , Algorithm 573: NL2SOL – An adaptive nonlinear least-squares algorithm, ACM

Trans. Math. Software, 7 (1981), pp. 369–383.
[79] J. E. Dennis, M. Heinkenschloss, and L. N. Vicente, Trust-region interior-point SQP

algorithms for a class of nonlinear programming problems, SIAM J. Control Optim., 36
(1998), pp. 1750–1794.
[80] J. E. Dennis and H. H. W. Mei, Two unconstrained optimization algorithms which use
function and gradient values, J. Optim. Theory Appl., 28 (1979), pp. 453–482.
[81] J. E. Dennis and J. J. Moré, A characterization of superlinear convergence and its

application to quasi-Newton methods, Math. Comp., 28 (1974), pp. 549–560.
[82] , Quasi-Newton methods, methods, motivation and theory, SIAM Rev., 19 (1977),
pp. 46–89.
[83] J. E. Dennis and R. B. Schnabel, Least change secant updates for quasi-Newton meth-
ods, SIAM Rev., 21 (1979), pp. 443–459.

166 BIBLIOGRAPHY
[84] , Numerical Methods for Unconstrained Optimization and Nonlinear Equations,

SIAM, Philadelphia, 1996.
[85] J. E. Dennis and V. Torczon, Direct search methods on parallel machines, SIAM J.
Optim., 1 (1991), pp. 448–474.
[86] J. E. Dennis and L. N. Vicente, Trust-region interior-point algorithms for minimization
problems with simple bounds, inApplied Mathematics and Parallel Computing, H. Fischer,
B. Riedmiller, and S. Schaffler, eds., Hidelberg, 1997, Springer, pp. 97–109.
[87] J. E. Dennis and H. F. Walker, Convergence theorems for least-change secant update
methods, SIAM J. Numer. Anal., 18 (1981), pp. 949–987.
[88] , Inaccuracy in quasi-Newton methods: Local improvement theorems, in Mathe-
matical Programming Study 22: Mathematical Programming at Oberwolfach II, North–
Holland, Amsterdam, 1984, pp. 70–85.
[89] J. E. Dennis and D. J. Woods, Optimization on microcomputers: The Nelder-Mead
simplex algorithm, in New Computing Environments: Microcomputers in Large-Scale
Scientific Computing, A. Wouk, ed., SIAM, Philadelphia, 1987, pp. 116–122.
[90] P. Deuflhard and V. Apostolescu, A Study of the Gauss-Newton Algorithm for the
Solution of Nonlinear Least Squares Problems, Tech. Rep. 51, Univ. Heidelberg, 1980,
preprint.
[91] P. Deuflhard, R. W. Freund, and A. Walter, Fast secant methods for the iterative
solution of large nonsymmetric linear systems, Impact Comput. Sci. Engrg., 2 (1990),
pp. 244–276.
[92] P. Deuflhard and G. Heindl, Affine invariant convergence theorems for Newton’s
method and extensions to related methods, SIAM J. Numer. Anal., 16 (1979), pp. 1–10.
[93] 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.
[94] J. C. Dunn, Global and asymptotic convergence rate estimates for a class of projected
gradient processes, SIAM J. Control Optim., 19 (1981), pp. 368–400.
[95] , On the convergence of projected gradient processes to singular critical points, J.
[96] , A projected Newton method for minimization problems with nonlinear inequality
constraints, Numer. Math., 53 (1988), pp. 377–409.
[97] J. C. Dunn and E. W. Sachs, The effect of perturbations on the convergence rates of
optimization algorithms, Appl. Math. Optim., 10 (1983), pp. 143–147.
[98] J. C. Dunn and T. Tian, Variants of the Kuhn–Tucker sufficient conditions in cones of
nonnegative functions, SIAM J. Control Optim., 30 (1992), pp. 1361–1384.
[99] 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.
[100] E. Eskow and R. B. Schnabel, Algorithm 695: Software for a new modified Cholesky
factorization, ACM Trans. Math. Software, 17 (1991), pp. 306–312.

BIBLIOGRAPHY 167
[101] A. V. Fiacco and G. P. McCormick, Nonlinear Programming: Sequential Unconstrained

Minimization Techniques, SIAM, Philadelphia, 1990.
[102] R. Fletcher, Generalized inverse methods for the best least squares solution of systems
of nonlinear equations, Comput. J., 10 (1968), pp. 392–399.
[103] , A new approach to variable metric methods, Comput. J., 13 (1970), pp. 317–322.
[104] , Practical Methods of Optimization, John Wiley and Sons, New York, 1987.
[105] R. Fletcher and M. J. D. Powell, A rapidly convergent descent method for minimiza-
tion, Comput. J., 6 (1963), pp. 163–168.
[106] R. Fletcher and C. M. Reeves, Function minimization by conjugate gradients, Comput.

J., 7 (1964), pp. 149–154.
[107] S. J. Fortune, D. M. Gay, B. W. Kernighan, O. Landron, R. A. Valenzuela, and

M. H. Wright, WISE design of indoor wireless systems, IEEE Computational Science
and Engineering, Spring (1995), pp. 58–68.
[108] J. Gablonsky, An Implemention of the Direct Algorithm, Tech. Rep. CRSC-TR98-29,

Center for Research in Scientific Computation, North Carolina State University, Raleigh,
August 1998.
[109] D. M. Gay, Computing optimal locally constrained steps, SIAM J. Sci. Statist. Comput.,
2 (1981), pp. 186–197.
[110] C. W. Gear, Numerical Initial Value Problems in Ordinary Differential Equations,

Prentice–Hall, Englewood Cliffs, NJ, 1971.
[111] 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.
[112] J. C. Gilbert and J. Nocedal, Global convergence properties of conjugate gradient

methods for optimization, SIAM J. Optim., 2 (1992), pp. 21–42.
[113] P. E. Gill and W. Murray, Newton-type methods for unconstrained and linearly con-
strained optimization, Math. Programming, 28 (1974), pp. 311–350.
[114] , Safeguarded Steplength Algorithms for Optimization Using Descent Methods,

Tech. Rep. NAC 37, National Physical Laboratory Report, Teddington, England, 1974.
[115] , Non-linear least squares and nonlinearly constrained optimization, in Numerical

Analysis, Lecture Notes in Mathematics 506, Springer-Verlag, Berlin, 1976.
[116] , Algorithms for the solution of the nonlinear least-squares problem, SIAM J. Numer.
Anal., 15 (1978), pp. 977–992.
[117] P. E. Gill, W. Murray, and M. H. Wright, Practical Optimization, Academic Press,

London, 1981.
[118] P. Gilmore, An Algorithm for Optimizing Functions with Multiple Minima, Ph.D. thesis,
North Carolina State University, Raleigh, 1993.
[119] , IFFCO: Implicit Filtering for Constrained Optimization, Tech. Rep. CRSC-TR93-
7, Center for Research in Scientific Computation, North Carolina State University, May
1993. Available by anonymous ftp from math.ncsu.edu in pub/kelley/iffco/ug.ps.

168 BIBLIOGRAPHY
[120] P. Gilmore and C. T. Kelley, An implicit filtering algorithm for optimization of functions
with many local minima, SIAM J. Optim., 5 (1995), pp. 269–285.
[121] P. A. Gilmore, S. S. Berger, R. F. Burr, and J. A. Burns, Automated optimization
techniques for phase change piezoelectric ink jet performance enhancement. in Proc.
IS&T’s NIP 13: 1997 International Conference on Digital Printing Technologies, Society
for Imaging Science and Technology, 1997, pp. 716–721.
[122] R. Glowinski, Numerical Methods for Nonlinear Variational Problems, Springer-Verlag,
New York, 1984.
[123] D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning,
Addison–Wesley, Reading, MA., 1989.
[124] D. Goldfarb, A family of variable metric methods derived by variational means, Math.
Comp., 24 (1970), pp. 23–26.
[125] A. A. Goldstein, Constructive Real Analysis, Harper and Row, New York, 1967.
[126] G. H. Golub and V. Pereyra, The differentiation of pseudo-inverses and nonlinear least
squares problems whose variables separate, SIAM J. Numer. Anal., 10 (1973), pp. 413–
432.
[127] G. H. Golub and C. G. Van Loan, Matrix Computations, Johns Hopkins University
Press, Baltimore, MD, 1983.
[128] A. Greenbaum, Iterative Methods for Solving Linear Systems, SIAM, Philadelphia, 1997.
[129] A. Griewank, On automatic differentiation, in Mathematical Programming: Recent De-
velopments and Applications, M. Iri and K. Tanabe, eds., Kluwer, Dordrecht, the Nether-
lands, 1989, pp. 83–108.
[130] A. Griewank and G. F. Corliss, eds., Automatic Differentiation of Algorithms: Theory,
Implementation, and Application, SIAM, Philadelphia, 1991.
[131] A. Griewank and P. L. Toint, Local convergence analysis for partitioned quasi-Newton
updates, Numer. Math., 39 (1982), pp. 429–448.
[132] , On the unconstrained optimization of partially separable functions, in Nonlinear
Optimization, M. J. D. Powell, ed., Academic Press, London, 1982.
[133] , Partitioned variable metric methods for large sparse optimization problems, Nu-
mer. Math., 39 (1982), pp. 119–137.
[134] L. Grippo and S. Lucidi, A globally convergent version of the Polak-Ribière conjugate
gradient method, Math. Programming, 78 (1997), pp. 375–392.
[135] W. A. Gruver and E. Sachs, Algorithmic Methods in Optimal Control, Pitman, London,
1980.
[136] W. Hackbusch, On the fast solving of parabolic boundary control problems, SIAM J.
Control Optim., 17 (1979), pp. 231–244.
[137] , Optimal H p,p/2 error estimates for a parabolic Galerkin method, SIAM J. Numer.
Anal., 18 (1981), pp. 681–692.
[138] , Multi-Grid Methods and Applications, Springer Ser. Comput. Math. 4, Springer-
Verlag, New York, 1985.

BIBLIOGRAPHY 169
[139] W. W. Hager, Rates of convergence for discrete approximations to unconstrained optimal

control problems, SIAM J. Numer. Anal., 13 (1976), pp. 449–472.
[140] M. Heinkenschloss, M. Ulbrich, and S. Ulbrich, Superlinear and Quadratic Conver-

gence of Affine Scaling Interior-Point Newton Methods for Problems with Simple Bounds
and Without Strict Complementarity Assumption, Tech. Rep. TR97-30, Department of
Computational and Applied Mathematics, Rice University, Houston, TX, December 1997.
[141] M. R. Hestenes and E. Steifel, Methods of conjugate gradient for solving linear sys-
tems, J. Res. Nat. Bureau Standards, 49 (1952), pp. 409–436.
[142] M. R. Hoare and P. Pal, Physical cluster mechanics: Statics and energy surfaces for
monoatomic systems, Adv. Phys., 20 (1971), pp. 161–196.
[143] J. H. Holland, Genetic algorithms and the optimal allocation of trials, SIAM J. Comput.,
2 (1973), pp. 88–105.
[144] , Adaption in Natural and Artificial Systems, University of Michigan Press, Ann
Arbor, MI, 1975.
[145] R. Hooke and T. A. Jeeves, “Direct search" solution of numerical and statistical prob-
lems, J. Assoc. Comput. Mach., 8 (1961), pp. 212–229.
[146] R. Horst, P. M. Pardolos, and N. V. Thoai, Introduction to Global Optimization,

Kluwer Academic Publishers, Dordrecht, the Netherlands, 1995.
[147] W. Huyer and A. Neumaier, Global Optimization by Multilevel Coordinate Search,

Institut für Mathematik, Universität Wien, 1997, preprint.
[148] D. M. Hwang and C. T. Kelley, Convergence of Broyden’s method in Banach spaces,

SIAM J. Optim., 2 (1992), pp. 505–532.
[149] G. Jasen, Investment dartboard: Pros and dart throwers both lose money, Your Money
Matters, Wall Street Journal, May 7, 1997.
[150] D. R. Jones, C. C. Perttunen, and B. E. Stuckman, Lipschitzian optimization without

the Lipschitz constant, J. Optim. Theory Appl., 79 (1993), pp. 157–181.
[151] L. Kantorovich and G. Akilov, Functional Analysis, 2nd ed., Pergamon Press, New
York, 1982.
[152] R. B. Kearfott, Rigorous Global Search: Continuous Problems, Kluwer, Dordrecht,

the Netherlands, 1966.
[153] C. T. Kelley, Identification of the support of nonsmoothness, in Large Scale Optimiza-

tion: State of the Art, W. W. Hager, D. W. Hearn, and P. Pardalos, eds., Kluwer Academic
Publishers B.V., Boston, 1994, pp. 192–205.
[154] , Iterative Methods for Linear and Nonlinear Equations, SIAM, Philadelphia, 1995.
[155] , Detection and remediation of stagnation in the Nelder-Mead algorithm using a

sufficient decrease condition, SIAM J. Optim., 10 (1999), pp. 43–55.
[156] C. T. Kelley, C. T. Miller, and M. D. Tocci, Termination of Newton/chord iterations

and the method of lines, SIAM J. Sci. Comput., 19 (1998), pp. 280–290.

170 BIBLIOGRAPHY
[157] C. T. Kelley and E. W. Sachs, Applications of quasi-Newton methods to pseu-

doparabolic control problems, in Optimal Control of Partial Differential Equations II—
Theory and Applications, Birkhäuser, Basel, 1987.
[158] , Quasi-Newton methods and unconstrained optimal control problems, SIAM J.

Control Optim., 25 (1987), pp. 1503–1516.
[159] , A pointwise quasi-Newton method for unconstrained optimal control problems,

Numer. Math., 55 (1989), pp. 159–176.
[160] , Pointwise Broyden methods, SIAM J. Optim., 3 (1993), pp. 423–441.
[161] , Multilevel algorithms for constrained compact fixed point problems, SIAM J. Sci.
Comput., 15 (1994), pp. 645–667.
[162] , Local convergence of the symmetric rank-one iteration, Computational Optimiza-

tion and Applications, 9 (1998), pp. 43–63.
[163] , A trust region method for parabolic boundary control problems, SIAM J. Optim.,
9 (1999), pp. 1064–1081.
[164] C. T. Kelley, E. W. Sachs, and B. Watson, A pointwise quasi-Newton method for

unconstrained optimal control problems, II, J. Optim. Theory Appl., 71 (1991), pp. 535–
547.
[165] H. F. Khalfan, R. H. Byrd, and R. B. Schnabel, A theoretical and experimental study

of the symmetric rank-one update, SIAM J. Optim., 3 (1993), pp. 1–24.
[166] S. Kirkpatrick, C. D. Geddat, and M. P. Vecchi, Optimization by simulated annealing,

Science, 220 (1983), pp. 671–680.
[167] J. R. Koehler and A. B. Owen, Computer experiments, in Handbook of Statistics, vol.

13, S. Shosh and C. R. Rao, eds., Elsevier, New York, 1996, pp. 261–308.
[168] J. Kostrowicki and L. Piela, Diffusion equation method of global minimization: Per-
formance for standard test functions, J. Optim. Theory Appl., 71 (1991), pp. 269–284.
[169] J. C. Lagarias, J. A. Reeds, M. H. Wright, and P. E. Wright, Convergence Properties

of the Nelder-Mead Simplex Algorithm in Low Dimensions, SIAM J. Optim., 9 (1998),
pp. 112–147.
[170] E. B. Lee and L. Markus, Foundations of Optimal Control Theory, John Wiley, New
York, London, Sydney, 1967.
[171] C. Lemaréchal, A view of line searches, in Optimization and Optimal Control, Auslander,
Oettli, and Stoer, eds., Lecture Notes in Control and Information Sciences 30, Springer-
Verlag, Berlin, 1981, pp. 59–78.
[172] K. Levenberg, A method for the solution of certain nonlinear problems in least squares,
Quart. Appl. Math., 4 (1944), pp. 164–168.
[173] R. M. Lewis and V. Torczon, Rank ordering and positive bases in pattern search algo-
rithms, Tech. Rep. 96-71, Institute for Computer Applications in Science and Engineering,
December, 1996.
[174] , Pattern search algorithms for linearly constrained minimization, SIAM J. Optim.,
9 (1999), pp. 1082–1099.

BIBLIOGRAPHY 171
[175] , Pattern search algorithms for linearly constrained minimization, SIAM J. Optim.,
10 (2000), pp. 917–941.
[176] D. C. Liu and J. Nocedal, On the limited memory BFGS method for large-scale opti-
mization, Math. Programming., 43 (1989), pp. 503–528.
[177] R. B. Long and W. C. Thacker, Data assimilation into a numerical equatorial ocean
model, part 2: Assimilation experiments, Dyn. Atmos. Oceans, 13 (1989), pp. 465–477.
[178] E. M. Lowndes, Vehicle Dynamics and Optimal Design, Ph.D. thesis, North Carolina
State University, Raleigh, 1998.
[179] S. Lucidi and M. Sciandrone, On the global convergence of derivative free methods
for unconstrained optimization, Università di Roma “La Sapienza”, Dipartimento di
Informatica e Sistemistica, 1997, reprint.
[180] D. G. Luenberger, Linear and Nonlinear Programming, Addison–Wesley, London,

1984.
[181] J. N. Lyness and C. B. Moler, Numerical differentiation of analytic functions, SIAM J.

Numer. Anal., 4 (1967), pp. 202–210.
[182] C. D. Maranas and C. A. Floudas, A global optimization method for Weber’s problem,
in Large Scale Optimization: State of the Art, W. W. Hager, D. W. Hearn, and P. Pardalos,
eds., Kluwer Academic Publishers B.V., Boston, 1994, pp. 259–293.
[183] D. W. Marquardt, An algorithm for least-squares estimation of nonlinear parameters,

J. Soc. Indust. Appl. Math., 11 (1963), pp. 431–441.
[184] J. M. Martinez, Quasi-Newton methods for solving underdetermined nonlinear simul-

taneous equations, J. Comput. Appl. Math., 34 (1991), pp. 171–190.
[185] E. S. Marwil, Exploiting Sparsity in Newton-Type Methods, Ph.D. thesis, Cornell Uni-
versity, Ithaca, NY, 1978.
[186] H. Matthies and G. Strang, The solution of nonlinear finite element equations, Internat.
J. Numer. Methods Engrg., 14 (1979), pp. 1613–1626.
[187] D. Q. Mayne and E. Polak, Nondifferential optimization via adaptive smoothing, J.

[188] K. I. M. McKinnon, Convergence of the Nelder-Mead Simplex Method to a Non-

Stationary Point, SIAM J. Optim., 9 (1998), pp. 148–158.
[189] E. H. Moore, General Analysis, Memoirs of the American Philosophy Society, 1935. I.
[190] J. J. Moré, The Levenberg-Marquardt algorithm: Implementation and theory, in Numer-

ical Analysis, G. A. Watson, ed., Lecture Notes in Mathematics 630, Springer-Verlag,
Berlin, 1977, pp. 105–116.
[191] , Trust regions and projected gradients, in System Modelling and Optimization,
Lecture Notes in Control and Information Sciences 113, Springer-Verlag, Berlin, 1988,
pp. 1–13.
[192] J. J. Moré and D. C. Sorensen, Computing a trust region step, SIAM J. Sci. Statist.
Comput., 4 (1983), pp. 553–572.

172 BIBLIOGRAPHY
[193] J. J. Moré and D. J. Thuente, Line search algorithms with guaranteed sufficient de-
crease, ACM Trans. Math. Software, 20 (1994), pp. 286–307.
[194] J. J. Moré and G. Toraldo, On the solution of large quadratic programming problems
with bound constraints, SIAM J. Optim., 1 (1991), pp. 93–113.
[195] J. J. Moré and S. J. Wright, Optimization Software Guide, SIAM, Philadelphia, 1993.
[196] J. J. Moré and Z. Wu, Global continuation for distance geometry problems, SIAM J.
Optim., 7 (1997), pp. 814–836.
[197] W. Murray and M. L. Overton, Steplength algorithms for minimizing a class of non-
differentiable functions, Computing, 23 (1979), pp. 309–331.
[198] S. G. Nash, Newton-type minimization via the Lanczos method, SIAM J. Numer. Anal.,
21 (1984), pp. 770–788.
[199] , Preconditioning of truncated-Newton methods, SIAM J. Sci. Statist. Comput., 6

(1985), pp. 599–616.
[200] L. Nazareth, A relationship between the BFGS and conjugate gradient algorithm and
its implications for new algorithms, SIAM J. Numer. Anal., 16 (1979), pp. 794–800.
[201] , Conjugate gradient methods less dependent on conjugacy, SIAM Rev., 28 (1986),
pp. 501–511.
[202] , A view of conjugate gradient-related algorithms for nonlinear optimization, in

Linear and Nonlinear Conjugate Gradient Methods, L. M. Adams and J. L. Nazareth,
eds., SIAM, Philadelphia, 1996, pp. 149–164.
[203] J. A. Nelder. Private Communication, 1998.
[204] J. A. Nelder and R. Mead, A simplex method for function minimization, Comput. J., 7
(1965), pp. 308–313.
[205] A. Neumaier, On convergence and restart conditions for a nonlinear conjugate gradient
method. Institut für Mathematik, Universität Wien, 1997, preprint.
[206] J. Nocedal, Updating quasi-Newton matrices with limited storage, Math. Comp., 35
(1980), pp. 773–782.
[207] , Theory of algorithms for unconstrained optimization, Acta Numerica, 1 (1991),

pp. 199–242.
[208] , Conjugate gradient methods and nonlinear optimization, in Linear and Nonlinear
Conjugate Gradient Methods, L. M. Adams and J. L. Nazareth, eds., SIAM, Philadelphia,
pp. 9–23.
[209] J. Nocedal and Y. Yuan, Combining Trust Region and Line Search Techniques, Tech.
Rep. OTC 98/04, Optimization Technology Center, Northwestern University, Chicago,
IL, 1998.
[210] J. A. Northby, Structure and binding of Lennard-Jones clusters: 13 ≤ n ≤ 147, J.

Chem. Phys., 87 (1987), pp. 6166–6177.
[211] J. M. Ortega and W. C. Rheinboldt, Iterative Solution of Nonlinear Equations in

Several Variables, Academic Press, New York, 1970.

BIBLIOGRAPHY 173
[212] R. Penrose, A generalized inverse for matrices, Proc. Cambridge Phil. Soc., 51 (1955),
pp. 406–413.
[213] L. R. Petzold, A description of DASSL: A differential/algebraic system solver, in Scien-

tific Computing, R. S. Stepleman et al., ed., North–Holland, Amsterdam, 1983, pp. 65–68.
[214] S. A. Piyawskii, An algorithm for finding the absolute extremum of a function, USSR
Comp. Math. and Math. Phys., 12 (1972), pp. 57–67.
[215] E. Polak and G. Ribière, Note sur la convergence de methodes de directions conjugées,
Rev Française Informat Recherche Operationelle, 3e Année, 16 (1969), pp. 35–43.
[216] B. T. Polyak, The conjugate gradient method in extremal problems, USSR Comp. Math.
and Math. Phys., 9 (1969), pp. 94–112.
[217] M. J. D. Powell, A FORTRAN Subroutine for Unconstrained Minimization, Requiring

First Derivatives of the Objective Function, Tech. Rep. AERE-R, 6469, Mathematics
Brance, A. E. R. E. Harwell, Berkshire, England, 1970.
[218] , A hybrid method for nonlinear equations, in Numerical Methods for Nonlinear
Algebraic Equations, P. Rabinowitz, ed., Gordon and Breach, New York, 1970, pp. 87–
114.
[219] , A new algorithm for unconstrained optimization, in Nonlinear Programming, J. B.

Rosen, O. L. Mangasarian, and K. Ritter, eds., Academic Press, New York, 1970, pp. 31–
65.
[220] , Convergence properties of a class of minimization algorithms, in Nonlinear Pro-

gramming 2, O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, eds., Academic Press,
New York, 1975, pp. 1–27.
[221] , Some global convergence properties of a variable metric algorithm without ex-
act line searches, in Nonlinear Programming, R. Cottle and C. Lemke, eds., American
Mathematical Society, Providence, RI, 1976, pp. 53–72.
[222] , Nonconvex minimization calculations and the conjugate gradient method, Lecture
Notes in Mathematics 1066, Springer-Verlag, Berlin, (1984), pp. 122–141.
[223] , On the global convergence of trust region algorithms for unconstrained minimiza-
tion, Math. Programming., 29 (1984), pp. 297–303.
[224] , Convergence properties of algorithms for nonlinear optimization, SIAM Rev., 28

(1986), pp. 487–500.
[225] , How bad are the BFGS and DFP methods when the objective function is quadratic,
Math. Programming., 34 (1986), pp. 34–47.
[226] , Update conjugate directions by the BFGS formula, Math. Programming, 38 (1987),
pp. 29–46.
[227] , Direct search algorithms for optimization calculations, Acta Numerica, 7 (1998),
pp. 287–336.
[228] K. Radhakrishnan and A. C. Hindmarsh, Description and Use of LSODE, the Liver-
more Solver for Ordinary Differential Equations, Tech. Rep. URCL-ID-113855, Lawrence
Livermore National Laboratory, Livermore, CA, December 1993.

174 BIBLIOGRAPHY
[229] W. Rudin, Principles of Mathematical Analysis, McGraw-Hill, New York, 1953.

[230] J. Sacks, W. J. Welch, T. J. Mitchell, and H. P. Wynn, Designs and analysis of
computer experiments, Statist. Sci., 4 (1989), pp. 409–435.
[231] R. B. Schnabel and E. Eskow, A new modified Cholesky factorization, SIAM J. Sci.
Statist. Comput., 11 (1990), pp. 1136–1158.
[232] G. A. Schultz, R. B. Schnabel, and R. H. Byrd, A family of trust-region-based algo-
rithms for unconstrained minimization with strong global convergence properties, SIAM
J. Numer. Anal., 22 (1985), pp. 47–67.
[233] V. E. Shamanskii, A modification of Newton’s method, Ukrain. Mat. Zh., 19 (1967),
pp. 133–138 (in Russian).
[234] L. F. Shampine, Implementation of implicit formulas for the solution of ODEs, SIAM J.
Sci. Statist. Comput., 1 (1980), pp. 103–118.
[235] , Numerical Solution of Ordinary Differential Equations, Chapman and Hall, New
York, 1994.
[236] L. F. Shampine and M. W. Reichelt, The MATLAB ODE suite, SIAM J. Sci. Comput.,
18 (1997), pp. 1–22.
[237] D. F. Shanno, Conditioning of quasi-Newton methods for function minimization, Math.
Comp., 24 (1970), pp. 647–657.
[238] , On the variable metric methods for sparse Hessians, Math. Comp., 34 (1980),
pp. 499–514.
[239] J. Sherman and W. J. Morrison, Adjustment of an inverse matrix corresponding 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.
[240] , Adjustment of an inverse matrix corresponding to a change in one element of a
given matrix, Ann. Math. Statist., 21 (1950), pp. 124–127.
[241] B. O. Shubert, A sequential method seeking the global maximum of a function, SIAM
J. Numer. Anal., 9 (1972), pp. 379–388.
[242] D. C. Sorensen, Newton’s method with a model trust region modification, SIAM J.
Numer. Anal., 19 (1982), pp. 409–426.
[243] , Minimization of a large-scale quadratic function subject to a spherical constraint,
SIAM J. Optim., 7 (1997), pp. 141–161.
[244] W. Spendley, G. R. Hext, and F. R. Himsworth, Sequential application of simplex
designs in optimisation and evolutionary operation, Technometrics, 4 (1962), pp. 441–
461.
[245] W. Squire and G. Trapp, Using complex variables to estimate derivatives of real func-
tions, SIAM Rev., 40 (1998), pp. 110–112.
[246] M. Srinivas and L. M. Patnaik, Genetic algorithms: A survey, Computer, 27 (1994),
pp. 17–27.
[247] T. Steihaug, The conjugate gradient method and trust regions in large scale optimization,
SIAM J. Numer. Anal., 20 (1983), pp. 626–637.

BIBLIOGRAPHY 175
[248] C. P. Stephens and W. Baritompa, Global optimization requires global information, J.

[249] G. W. Stewart, Introduction to Matrix Computations, Academic Press, New York, 1973.
[250] D. Stoneking, G. Bilbro, R. Trew, P. Gilmore, and C. T. Kelley, Yield optimiza-

tion using a GaAs process simulator coupled to a physical device model, IEEE Trans.
Microwave Theory and Techniques, 40 (1992), pp. 1353–1363.
[251] D. E. Stoneking, G. L. Bilbro, R. J. Trew, P. Gilmore, and C. T. Kelley, Yield

optimization using a GaAs process simulator coupled to a physical device model, in Proc.
IEEE/Cornell Conference on Advanced Concepts in High Speed Devices and Circuits,
IEEE, Piscataway, NJ, 1991, pp. 374–383.
[252] K. Tababe, Continuous Newton-Raphson method for solving an underdetermined system

of nonlinear equations, J. Nonlinear Anal., Theory Methods Appl., 3 (1979), pp. 495–503.
[253] , Global analysis of continuous analogs of the Levenberg-Marquardt and Newton-

Raphson methods for solving nonlinear equations, Ann. Inst. Statist. Math., B, 37 (1985),
pp. 189–203.
[254] T. Tian and J. C. Dunn, On the gradient projection method for optimal control problems
with nonnegative L2 inputs, SIAM J. Control Optim., 32 (1994), pp. 516–537.
[255] P. L. Toint, On sparse and symmetric matrix updating subject to a linear equation, Math.
Comp., 31 (1977), pp. 954–961.
[256] , On the superlinear convergence of an algorithm for solving a sparse minimization

problem, SIAM J. Numer. Anal., 16 (1979), pp. 1036–1045.
[257] , Towards an efficient sparsity exploiting Newton method for minimization, in Sparse
Matrices and Their Uses, I. S. Duff, ed., Academic Press, New York, 1981, pp. 57–88.
[258] , On large scale nonlinear least squares calculations, SIAM J. Sci. Statist. Comput.,
8 (1987), pp. 416–435.
[259] , Global convergence of a class of trust–region methods for nonconvex minimization

in Hilbert space, IMA J. Numer. Anal., 8 (1988), pp. 231–252.
[260] V. Torczon. Private communication, 1997.
[261] , Multidirectional Search, Ph.D. thesis, Rice University, Houston, TX, 1989.
[262] , On the convergence of the multidimensional search algorithm, SIAM J. Optim., 1

(1991), pp. 123–145.
[263] , On the convergence of pattern search algorithms, SIAM J. Optim., 7 (1997),

pp. 1–25.
[264] L. N. Trefethen and D. Bau, Numerical Linear Algebra, SIAM, Philadelphia, 1996.
[265] P. van Laarhoven and E. Aarts, Simulated Annealing, Theory and Practice, Kluwer,
Dordrecht, the Netherlands, 1987.
[266] L. N. Vicente, Trust-Region Interior-Point Algorithms for a Class of Nonlinear Program-

ming Problems, Ph.D. thesis, Rice University, Houston, TX, 1996.

176 BIBLIOGRAPHY
[267] F. H. Walters, L. R. Parker, S. L. Morgan, and S. N. Demming, Sequential Simplex

Optimization, CRC Press, Boca Raton, FL, 1991.
[268] B. Watson, Quasi-Newton Methoden für Minimierungsprobleme mit strukturierter

Hesse-Matrix, Diploma Thesis, Universität Trier, 1990.
[269] J. Werner, Über die globale konvergenz von Variable-Metric Verfahren mit nichtexakter
Schrittweitenbestimmung, Numer. Math., 31 (1978), pp. 321–334.
[270] T. A. Winslow, R. J. Trew, P. Gilmore, and C. T. Kelley, Doping profiles for optimum
class B performance of GaAs mesfet amplifiers, in Proc. IEEE/Cornell Conference on
Advanced Concepts in High Speed Devices and Circuits, IEEE, Piscataway, NJ, 1991,
pp. 188–197.
[271] , Simulated performance optimization of GaAs MESFET amplifiers, in Proc.

IEEE/Cornell Conference on Advanced Concepts in High Speed Devices and Circuits,
IEEE, Piscataway, NJ, 1991, pp. 393–402.
[272] P. Wolfe, Convergence conditions for ascent methods, SIAM Rev., 11 (1969), pp. 226–
235.
[273] , Convergence conditions for ascent methods II: Some corrections, SIAM Rev., 13
(1971), pp. 185–188.
[274] M. H. Wright, Direct search methods: Once scorned, now respectable, in Numerical
Analysis 1995, Proc. 1995 Dundee Bienneal Conference in Numerical Analysis, D. F.
Griffiths and G.A. Watson, eds., 1996, Addison–Wesley Longman, Harlow, U.K., pp. 191–
208.
[275] S. J. Wright, Compact storage of Broyden-class quasi-Newton matrices, Argonne Na-
tional Laboratory, Argone, IL, 1994, preprint.
[276] S. J. Wright and J. N. Holt, An inexact Levenbert-Marquardt method for large sparse
nonlinear least squares, J. Austral. Math. Soc. Ser. B, 26 (1985), pp. 387–403.
[277] Z. Wu, The effective energy transformation scheme as a special continuation approach
to global optimization with application to molecular conformation, SIAM J. Optim., 6
(1996), pp. 748–768.
[278] T. J. Ypma, The effect of rounding errors on Newton-like methods, IMA J. Numer. Anal.,
3 (1983), pp. 109–118.
[279] S. K. Zavriev, On the global optimization properties of finite-difference local descent

algorithms, J. Global Optim., 3 (1993), pp. 67–78.
[280] C. Zhu, R. H. Byrd, P. Lu, and J. Nocedal, L-BFGS-B—FORTRAN subroutines for

large-scale bound constrained optimization, ACM Trans. Math. Software, 23 (1997),
pp. 550–560.

Index
-active set, 97 chord method, 19

-inactive set, 97 Newton’s method, 15
Convex quadratic, 6, 45
Actual reduction, 40 Coordinate descent algorithm, 146
trust region, 51 Critical point, 5
Adjoint equation, 10
Adjoint variable, 10 DACE, 149
Armijo rule, 39, 78 Default choice, 82
Dennis–Moré condition, 76
Backtracking, 39 Descent direction, 40
BFGS DFP update, 81
algorithm, 80 Difference approximation
global convergence, 78 Hessian, 20
implementation, 78 Hessian–vector product, 20
limited memory, 79 DIRECT
local convergence, 72 method, 135
modified update, 80 Direct search algorithms, 135
restarted, 79 Direction of negative curvature, 9, 30
update, 71 Dogleg, 58
Bounded deterioration, 74 classical, 58
Breakdown
conjugate gradient, 7 Feasible
point, 87
Cauchy decrease, 62 set, 87
Cauchy point, 55 Fletcher–Reeves formula, 49
CG-Trust region optimizer, 65 Forcing term, 28
CGTRUST Forward difference CG
algorithm, 65 algorithm, 30
Cholesky factorization, 7, 16 Forward difference PCG
Chord method, 19 algorithm, 31
algorithm, 19 Frobenius norm, 77
Classical dogleg Fundamental theorem of calculus, 4
algorithm, 63
Conjugate gradient, 7, 28 Gauss–Newton
algorithm, 7 damped, 47
Constrained optimization problem, 3 Gauss–Newton method, 23
Constraints, 87 Genetic algorithms, 112
active, 87 Global convergence
inactive, 87 Armijo rule, 43
Control variable, 10 BFGS–Armijo, 78
Convergence dogleg trust region, 52
r-type, 14 gradient projection method, 95
Convergence theorem projected BFGS method, 104
177
178 INDEX
scaled srojected gradient, 99 inexact Newton method, 29, 30

Global minimization problem, 3 Newton’s method, 16
Global minimizer, 3 Projected BFGS method, 104
Globally convergent algorithm, 39 Projected Newton method, 101
Gradient of f , 4 q-type, 13
Gradient Projection Local improvement, 18
algorithm, 91 Local linear model
Gradient projection, 91 for equations, 14
Local minimizer, 3
Hessian of f , 4 nondegenerate, 90
Hooke–Jeeves Local quadratic model
algorithm, 135, 145 exact Hessians, 15
convergence, 148
exploratory move, 145 MDS, 135, 143
pattern move, 146 convergence, 145
Measure of stationarity, 93
Implicit filtering
Minimizer
restart, 127
global, 3
scales, 124
local, 3
Indefinite matrix, 4
Minimum, 3
Indices
Minimum at all scales, 127
active, 87
Minimum norm solution, 23, 25
inactive, 87
Model Hessian, 40
Inexact Newton
Moore–Penrose inverse, 26
method, 28
Multidirectional search, 135, 143
step, 28
convergence, 145
Initial
guess, 4
Necessary condition
iterate, 4
first-order, 5
Inner iteration, 28
Necessary conditions
Interval methods, 112
bound constraints
Krylov subspace, 8 one variable, 88
first-order
Large residual problem, 22 bound constraints, 88
Lennard–Jones function, 120 for optimality, 5
Levenberg–Marquardt Negative curvature, 9, 30, 36
method, 47 Nelder–Mead
parameter, 47, 56 oriented restart, 141
Line search, 39 simplex method, 135
actual reduction, 40 stagnation
exact, 41, 86 detection, 141
predicted reduction, 40 Newtcg
Line search algorithm, 40 algorithm, 32
Linear model, 14, 40 Newton step, 15
Lipschitz Newton’s method, 14
constant, 14 algorithm, 16
continuity, 14 Newton-iterative methods, 28
Local convergence Nondegenerate local minimizer, 90
BFGS, 72 Nonlinear conjugate gradient
chord method, 19 algorithm, 48
implicit filtering, 124 convergence, 49

INDEX 179
Fletcher–Reeves, 49 Secant method

Polak–Ribière, 49 classical, 71
Nonlinear least squares, 11, 22 Sensitivity equations, 11, 12
Jacobian, 22 Shamanskii method, 20
overdetermined, 22 Sherman–Morrison
residual, 22 formula, 81
underdetermined, 22 Shubert algorithm, 149
Ntrust Simple decrease, 40
algorithm, 63 Simplex
condition, 112
Objective function, 3 convex hull, 112
Optimality definition, 112
necessary conditions, 5 diameter, 112
sufficient conditions, 6 directions, 112
Optimization problem gradient
constrained, 3 forward, 113
easiest, 12 nonsingular, 112
unconstrained, 3 reflected, 115
Oriented lengths, 112 vertex, 112
Outer iteration, 28 Simplex gradient
central, 115
Parameter identification, 11 Simulated annealing, 112
Polak–Ribière formula, 49 Singular value decomposition, 25
Positive definite matrix, 4
Singular values, 26
Positive semidefinite matrix, 4
Singular vectors, 26
Preconditioned CG
Small residual problem, 22
algorithm, 8
spd, 4
Preconditioning
Spendley–Hext–Himsworth algorithm, 159
conjugate gradient, 8
SR1 update, 81
Predicted reduction
Stagnation, 19
line search, 40
Nelder–Mead
trust region, 51
example, 139
Projected BFGS method, 104
Standard assumptions, 14
Projected Newton method, 100
State equation, 10
algorithm, 100
State variable, 10
Projection
Stationary point, 5, 43, 100
onto Ω, 89
bound constraints, 88
PSB update, 81
nondegenerate, 90
Quadratic model, 6, 40 Steepest descent
Quadratic objective functions, 6 direction, 39
Quasi-Newton methods, 71 method, 39
Stencil failure, 117, 123
Reduced Hessian Steplength, 39
definition, 89 Stiff initial value problems, 11
Response surfaces, 148 Strict complementarity, 90
Sufficient conditions
Safeguarding, 44 for optimality, 6
Scaled gradient projection, 99 Sufficient decrease, 39, 41
Scaling, 46 gradient projection, 91
Scaling matrix, 99 projected Newton, 97
Secant equation, 71 simplex form, 138

180 INDEX
Surrogate models, 148
Taylor’s theorem, 5
Termination
bound constrained problems, 91
small function differences, 21
small gradients, 21
small steps, 21
unconstrained problems, 21
Tosca, Floria, xiv
TR-CG
algorithm, 64
Truncated Newton methods, 28
Trust region, 50
actual reduction, 51
algorithm
adjustment, 51
paradigm, 52
dogleg, 58
classical, 58
convergence theorem, 52
Levenberg–Marquardt
algorithm, 58
parameter, 57
methods, 50
Newton point, 58
predicted reduction, 51
problem, 50
radius, 50
trial solution, 50
trial step, 50
unidirectional, 54
Trust region CG
algorithm, 64
Unconstrained optimization problem, 3
Weber’s problem, 118, 127, 152

Wolfe conditions, 49
strong, 49
Youngman
Henny, 159
Zero residual problem, 22

Kelley - Iterative Methods For Optimization-SIAM (1999) PDF

Uploaded by

Copyright:

Available Formats

Kelley - Iterative Methods For Optimization-SIAM (1999) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Kelley - Iterative Methods For Optimization-SIAM (1999) PDF

Uploaded by

Copyright:

Available Formats

Copyright ©1999 by the Society for Industrial and Applied Mathematics.

How to Get the Software xv

I Optimization of Smooth Functions 1

2 Local Convergence of Newton’s Method 13

4 The BFGS Method 71

5 Simple Bound Constraints 87

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

5.8 Exercises on Bound Constrained Optimization . . . . . . . . . . . . . . . . . . 108

II Optimization of Noisy Functions 109

7 Implicit Filtering 123

8 Direct Search Algorithms 135

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

653-1415, [email protected], http://www.mathworks.com.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

How to Get the Software

or from SIAM’s World Wide Web server at

One can obtain MATLAB from

Optimization of Smooth Functions

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

1.1 The Problem

(1.1) f (x∗ ) ≤ f (x) for all x near x∗ .

It is standard to express this problem as

(1.2) min f (x)

(1.3) f (x∗ ) ≤ f (x) for all x,

(1.4) f (x∗ ) ≤ f (x) for all x ∈ U near x∗ .

Similar to (1.2) we express this as

4 ITERATIVE METHODS FOR OPTIMIZATION

In optimization definiteness or semidefiniteness of the Hessian plays an important role in

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

(1.7) f (x∗ + e) = f (x∗ ) + ∇f (x∗ )T e + eT ∇2 f (x∗ )e/2 + o(

1.3 Necessary Conditions

Setting ∇f (x∗ ) = 0, dividing by t, and setting t = 0 in (1.8), we obtain

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

6 ITERATIVE METHODS FOR OPTIMIZATION

1.4 Sufficient Conditions

Theorem 1.4.1. Let f be twice continuously differentiable in a neighborhood of x∗ . Assume

Proof. Let 0 = u ∈ RN . For sufficiently small t we have

Hence, if λ > 0 is the smallest eigenvalue of ∇2 f (x∗ ) we have

for t sufficiently small. Hence x∗ is a local minimizer.

1.5 Quadratic Objective Functions

We may, without loss of generality, assume that H is symmetric because

for all x. The symmetry of H implies that

Deﬁnition 1.5.1. The quadratic function f in (1.9) is convex if H is positive semidefinite.

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

1.5.1 Positive Definite Hessian

or if too many iterations have been taken.

f (xk ) ≤ f (x) for all x ∈ x0 + Kk ,

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

8 ITERATIVE METHODS FOR OPTIMIZATION

where Kk is the Krylov subspace

Kk = span(r0 , Hr0 , . . . , H k−1 r0 )

for an spd matrix H. κ(H) is the l2 condition number

κ(M 1/2 HM 1/2 )

(1.13) M 1/2 HM 1/2 z = M 1/2 b

Buy this book from SIAM at http://www.ec-securehost.com/SIAM/FR18.html.

(c) if k = 1 then β = 0 and p = z

1.5.2 Indefinite Hessian

Proof. Let 0 = u ∈ RN . For sufficiently small t we have

(1.21) u + cu + ku = 0; u(0) = u0 , u (0) = 0,

w1 + u + cw1 + kw1 = 0; w1 (0) = w1 (0) = 0,

u + cu + ku = sin(ωt + φ); u(0) = 10, u (0) = 0.

Mc (x) = F (xc ) + F (xc )(x − xc ).

(2.2) x+ = xc − F (xc )−1 F (xc ).

x+ = xc − (∇2 f (xc ) + ∆(xc ))−1 (∇f (xc ) + g (xc ))

The convergence theory follows from Theorem 2.3.4 with g = 0 and ∆ = O(