Quasi Linear Ization

Quasilinearization and Nonlinear
Boundary-Value Problems
Richard E. ,Bellman and Robert E. Kalaba
The RAND Corporation
AMERICAN ELSEVIER PUBLISHING COMPANY, INC
New York 196 S
SOLE DISTRIBUTORS FOR UREAT BRITAIN
ELSEVIER PUBLISHING COMPANY, LTD.
Barking, Essex, England
SOLE DISTRIBUTORS FOR THE CONTINENT OF EUROPE
ELSEVIER PUBLISHING COMPANY
Amsterdam, The Netherlands
Library of Congress Catalog Card Number: 65-22807
COPYRIGHT 1965 BY THE RAND CORPORATION
ALL RIGHTS RESERVED. THIS BOOK OR ANY PART THEREOF
MUST NOT BE REPRODUCED IN ANY FORM WITHOUT THE WRITTEN
PERMISSION OF THE PUBLISHER, AMERICAN ELSEVIER PUBLISHING
COMPANY, INC., 52 VANDERBILT AVENUE, NEW YORK 10017, N.Y.
PUNTED IN THE UNITED STATES
,I
Introduction
CONTENTS
CHAPTER ONE
The Riccati Equation
1
1. Introduction. 7
2. Newton-Raphson Method 7
3. Multidimensional Version. 11
4. Square Roots 12
5. Quasilinearization . 13
6. The Riccati Equation 14
7. The First Order Linear Equation 15
8. Solution of Riccati Equation in Terms of Maximum Operation 15
9. Upper and Lower Bounds 18
10. Successive Approximations via Quasilinearization 19
11. Monotonicity 20
12. Discussion 21
13. A Fundamental Lemma 22
Comments and Bibliography 22
CHAPTER TWO
Two-Point Boundary-Value Problems for
Second-Order Differential Equations
1. Introduction. 27
2. Two-Point Boundary-Value Problems for Second-Order Linear
Differential Equation . 28
3. The Inhomogeneous Equation 29
VI CONT.NTS
4. Vector-Matrix Approach 32
5. Green's Functions 34
6. Convexity 34
7. Quasilinearization 36
8. Discussion 37
9. Existence and Boundedness 37
10. Convergence 38
11. Convergence of the Picard Algorithm 40
12. A Numerical Example 41
13. General Second-Order Nonlinear Differential Equations 42
15. Calculus of Variations 44
16. Quasilinearization . 46
17. Upper and Lower Bounds via Quasilinearization 48
18. Solution of Inner Problem 50
19. Further Uses of Quasilinearization 50
20. Dynamic Programming 51
21. Invariant Imbedding 52
22. Combination of Dynamic Programming and Quasilineariza-
tion . 53
CHAPTER THREE
Monotone Behavior and Differential Inequalities
1. Monotonicity 57
2. An Elementary Approach 58
3. Relations Between Solutions and Coefficients. 59
4. A Useful Factorization of the Second-Order Linear Differ-
ential Operator . 60
5. A Positivity Result 62
6. A Related Parabolic Partial Differential Equation 63
7. Characteristic Values Again 66
8. Variational Approach . 67
9. Discussion 68
10. Improvement of Convergence Estimates 70
11. Discussion 71
CONTENTS
CHAPTER FOUR
Systems of Differential Equations, Storage and
Differential Approximation
VII
1. Introduction. 75
2. Quasilinearization Applied to Systems 76
3. Solution of Linear System 78
5. Multipoint Boundary-Value Problems 80
6. Successive Approximations and Storage 80
7. Simultaneous Calculation of Approximations 81
8. Discussion . 83
9. Back and Forth Integration . 85
10 .. Differential-Difference Equations 85
11. Reduction to Differential Equations. 86
12. Functional Differential Equations 87
13. Differential Approximation 88
14. Simple Version of Differential Approximation 89
15. A Renewal Equation 91
16. Discussion 93
17. Storage and Memory 95
18. Monotonicity for Systems 95
19. Monotonicity for Nth-Order Linear Differential Equations. 97
20. Discussion 98
21. Use of the Gram-Schmidt Orthogonalization Procedure in
the Numerical Solution of Linear Boundary-Value Problems 98
22. Dynamic Programming 103
23. Invariant Imbedding 104
CHAPTER FIVE
Partial Differential Equations
1. Introduction.
2. Parabolic Equations
3. A Specific Equation
4. Difference Approximation
5. Comparison Between Picard and Quasilinearization
6. Gradient Techniques .
111
112
112
113
114
114
vIIi
7. An Elliptic Equation .
8. Computational Aspects
9. Positivity and Monotonicity
10. The Hopf-Lax Equation
Comments and Bibliography
CHAPTER SIX
CONTINTS
116
117
121
122
123
Applications in Physics, Engineering, and Biology
1. Introduction 125
2. Optimal Design and Control 125
3. An Example 127
4. Numerical Results . 129
5. An Inverse Problem in Radiative Transfer 130
6. Analytic Formulation . 132
8. The Van der Pol Equation 134
9. Orbit Determination as a Multipoint Boundary-Value
Problem . 137
11. Periodogram Analysis 139
13. Discussion 144
14. Periodic Forcing Terms 144
15. Estimation of Heart Parameters. 144
16. Basic Assumptions. 146
17. The Inverse Problem . 147
18. Numerical Experiments 148
CHAPTER SEVEN
Dynamic Programming and Quasilinearization
1. Introduction .
2. A Basic Functional Equation
3. Approximation in Policy Space
4. Dynamic Programming and Quasilinearization
5. Functional Equations
6. Discussion
153
153
154
156
157
158
CONTENTS ix
7. Dynamic Programming and Differential Approximation. 158
8. Dynamic Programming and the Identification of Systems 159
9. Discussion 161
APPENDICES
One: Minimum Time Program 163
Two: Design and Control Program . 167
Three: Program for Radiative Transfer-Inverse Problem for Two
Slabs, Three Constants 183
Four: Van Der Pol Quation Program 185
Five: Orbit Determination Program 189
Six: Cardiology Program 193
Index 203
Quasilinearization and Nonlinear
Boundary-Value Problems
INTRODUCTION
Modern electronic computers .can provide the numerical solution
of systems of one thousand simultaneous nonlinear ordinary differ-
ential equations, given a complete set of initial conditions, with
accuracy and speed. It follows that as soon as physical, economic,
engineering, and biological problems are reduced to the task of
solving initial value problems for ordinary differential equations,
they are well on their way to complete solution. Hence, the great
challenge to the mathematician is to find ways to construct the
requisite theories and formulations, approximations and reductions.
Sometimes, as in planetary theory, the basic equations are ordinary
differential equations. In this case, however, some of the fundamental
problems do not provide us with a full set of initial conditions. In
fact, the basic problem of orbit determination is to determine a com-
plete set of initial conditions, given certain angular measurements
made at various times. In other areas, particularly in mathematical
physics and mathematical biology, the underlying equations are
partial differential equations, differential-integral equations, differ-
ential-difference equations, and equations of even more complex
type. Many preliminary analyses of the precise questions to be asked,
and the analytical procedures to be used, must be made in order to
exploit fully the powerful computational capabilities now available.
Whereas in earlier times the preferred procedure was to simplify so
as to obtain linear functional equations, the current aim is to trans-
form all computational questions to initial value problems for
ordinary differential equations, be they linear or nonlinear. The
theories of dynamic programming and invariant imbedding achieve
this in a number of areas through the introduction of new state
2 QUASILINEARIZATION AND NONLINEAR BOUNDARYVALUE PROBLEMS
variables and the use of semigroup properties in space, time, and
structure. Quasilinearization achieves this objective by com bini ng
linear approximation techniques with the capabilities of the digital
computer in various adroit fashions. The approximations are care-
fully constructed to yield rapid convergence, and monotonicity as
well, in many cases.
The origin of quasilinearization lies in the theory of dynamic
programming. The connections with classical geometry and analysis
are, nonetheless, quite diverse and do not constitute any simple
pattern. A basic ingredient is geometric duality theory of the type
associated with Minkowski, Fenchel, Kuhn and Tucker, and many
others; and there is a close association with the duality introduced
into the calculus of variations by Friedrichs and since extended by
so many others. In another direction, there is a strong heritage from
the theory of differential inequalities started by Caplygin and con-
tinued by a very active Russian school. Intimately connected with
this is the modern theory of positive and monotone operators
created by Collatz and vigorously developed by Redheffer and others.
This, in turn, is associated with the theory of generalized convexity
pursued by Beckenbach, Bonsall, Peixoto, and a number of other
analysts. As far as many applications are concerned, we seem merely
to be applying Newton-Raphson-Kantorovich approximation tech-
niques in function space.
Without dwelling at length on any of these very intriguing areas,
which themselves merit monographs, we mention the connections
in the text and give the interested reader a number of references for
further reading.
The objectives of'the theory of quasilinearization are easily stated.
First, we desire a uniform approach to the study of the existence and
uniqueness of the solutions of ordinary and partial differential
equations subject to initial and boundary-value conditions. Further-
more, we want representation theorems for these solutions in terms
of the solutions of linear equations. Finally, we want a uniform
approach to the numerical solution of both descriptive and variational
INTRODUCTION 3
problems which possesses various monotonicity properties and
rapidity of convergence.
Throughout this work, when we use the terms "computational
solution" or "numerical solution," we are thinking in terms of
algorithms to be carried out by means of digital computers.
It is esthetically pleasing, pedagogically desirable, and efficient
scientifically that the same conceptual and analytic approach serves
all of these purposes. This volume is intended as an introduction to
quasilinearization, aimed both at those who are solely interested in
the analysis, and those who are primarily busied with applications.
Let us now turn to a brief description of the contents of the seven
chapters. To acquaint the reader with the analytic techniques in their
simplest form, we begin in Chapter One with a discussion of the
application of quasilinearization to the study of the Riccati equation,
(1)
This is one of the most important ordinary differential equations in
analysis, both in its own right, and because of its connection with the
second-order linear differential equation
(2)
w" + a
1
(t)w' + alt)w = o.
It is shown how quasilinearization permits us to obtain an explicit
analytic representation for u in terms of quadratures and the maxi-
mum operation. From this, both upper and lower bounds for u and
w can be deduced.
In Chapter Two, we turn to a discussion of two-point boundary-
value problems of the form
(3) U" = g(u',lI,t),
with the boundary conditions
(4) h
1
(u'(0),u(0)) = 0, hlu'(l),u(l)) = o.
The quadratic convergence of the sequence of approximations
generated by quasilinearization is studied, and some numerical
examples are given.
.. QUASI LINEARIZATION AND NONLINEAR BOUNDARYVALUE PROBLEMS
Chapter Three is devoted to the examination of conditions under
which the sequence of approximations converges monotonically.
We are led to study linear differential inequalities of the form
(5)
u" + p(t)U' + q(t)U ~ O.
In this way, we encounter Green's functions. To treat the sign-
preserving character of these functions, we employ a number of
interesting techniques, such as factorization of the operator, varia-
tional methods, and difference approximations to partial differential
equations. These questions are intimately related to characteristic
value problems.
In Chapter Four, some corresponding problems of monotonicity
are treated for higher-order differential equations and for linear
systems. The analytic results here are quite sketchy and not easy
to obtain. A number of examples of the numerical efficacy of the
methods are given.
A close scrutiny of the effectiveness of the numerical procedures
suggested shows that the use of successive approximations in a
routine fashion can result in storage difficulties when the order of
the differential equation or system is high. We are prompted then to
develop techniques for using successive approximations without
excessive reliance on rapid-access storage, and, in general, to study
the fascinating area of the storage and retrieval of information. A
very simple method for avoiding storage is presented, and then applied
to the numerical solution of differential-difference equations of the
form
(6) u'( t) = g( u( t), u( t - t
1
), t),
and to functional-difference equations of the form
(7) u'(t) = g(u(t),u(h(t)),t).
Equations of this type play important roles in modern mathematical
biology.
In a further effort to overcome the storage problem, we turn to the
method of differential approximation. This method was developed
INTRODUCTION
5
in collaboration with J. M. Richardson of the Hughes Research
Laboratories at Malibu to provide analytic approximations to various
classes of complex functional equations. A general formulation for
differential equations is the following.
Given a vector equation of the form
(8)
dx
dt = g(x,t),
x(O) = c,
we wish to find a more tractable equation of the form
dy
- = h(y,t,a),
dt
yeO) = b, (9)
with the property that y '"'-' x.
Here a and b are vectors to be chosen so that IIx - yll, a suitably
defined measure, such as J: (x - y, x - y) dt, is small. This
variational problem is amenable to the technique of quasilinear-
ization. As an application of this approach, we show how to obtain
an accurate solution of a renewal equation such as
(to)
r
t
2
U(t) = 1 + Jo e-<t-s) u(s) ds
by means of a finite system of ordinary differential equations.
In Chapter Five, we present the results of some applications of
quasilinearization to the solution of the nonlinear elliptic partial
differential equation
(11 )
and the nonlinear parabolic partial differential equation
(12) u
t
= u"'''' + g(u),
subject to various initial and boundary conditions.
Although only a few cases have been run, because of the time
involved, there are clear indications that there is much to be gained
in both time and accuracy by employing quasilinearization in the
numerical solution of partial differential equations. Much, however,
remains to be done before any definitive statements can be made
6 QUASILINEARIZATION AND NONLINEAR BOUNDARY-VALUE PROBLEMS
comparable to those that can be uttered as far as ordinary differential
equations are concerned.
Chapter Six contains a number of examples which illustrate some
of the ways in which quasilinearization can be applied to treat
descriptive and variational processes arising in design and control,
in orbit determination, in the detection of periodicities, and in the
identification of systems. This last topic covers inverse problems in
radiative transfer and cardiology.
The final chapter on dynamic programming begins with an ex-
planation of the way in which approximation in policy space leads
to quasilinearization, continues with an indication of how dynamic
programming can be used to bypass two-point boundary values
arising in quadratic variational problems, and concludes with a
discussion of how local quasilinearization and global dynamic
programming can be used to widen the domains of applicability of
both theories.
As always, we have imposed upon friends and colleagues for their
advice and opinions. We particularly wish to thank T. A. Brown,
O. Gross, M. Juncosa, and R. Sridhar in this connection. Without
the aid of J. Buell, B. Kotkin, and especially H. Kagiwada in writing
the computer programs, and in pursuing the minutiae that make the
difference between success and failure, much of this would have been
quite visionary. Finally, we express our appreciation of Jeanette
Blood whose patience, industry, intelligence, and sense of humor
enabled us to complete the chore of book writing.
All that we have done is a prologue to what remains to be done
in the future. We hope that others will find research in these domains
as interesting and rewarding as we have, and we hope also that our
methods will be used to cultivate intensively much that we have only
sampled.
This study was prepared as part of a program of research under-
taken by The RAND Corporation for the Advanced Research Pro-
jects Agency, the National [nstitutes of Health, and United States
Air Force Project RAND.
Chapter One
THE RICCATI EQUATION
I. INTRODUCTION
In this chapter we wish to apply approximation techniques to the
study of one of the fundamental equations of mathematical analysis,
the first-order nonlinear differential equation
(1.1) u' + u
2
+ p(t)u + q(t) = 0,
known as the Riccati equation. This equation plays a fundamental
role in the analysis of the behavior of the solutions of the second-
order linear differential equation
(1.2) w" + p(t)w' + q(t)w = 0,
and thus occupies an important position in quantum mechanics in
connection with the study of the Schrodinger equation. Furthermore,
together with its multidimensional analogues, it enjoys a central
place in dynamic programming and invariant imbedding.
Consideration of the equation of (1.1) will give us an opportunity
to illustrate the use of quasilinearization techniques, untrammeled
by arithmetic and algebraic underbrush. The basic properties of
monotonicity and quadratic convergence can be established in
particularly simple fashion, and we can thus, in a leisurely fashion,
set the pattern for much that follows.
2. NEWTON-RAPHSON METHOD
To motivate an approach we shall use repeatedly throughout this
volume, let us begin with the problem of finding a sequence of
approximations to the root of the scalar equation
(2.1 ) f( x) = 0.
7
y
x
Fig. 1.1
We shall assume that f(x) is monotone decreasing for all x and
strictly convex, that is,r(x) > O. (See Fig. 1.1.) Hence, the root r is
simple, /,(r) =;1= O.
Let Xo be an initial approximation to the root r, with Xo < r
(j(xo) > 0), and suppose that we approximate to f(x) by a linear
y
Fig. 1.1
THE RICCATI EQUATION 9
function of x determined by the value and slope of the function j(x)
at x = Xo,
(2.2)
(See Fig. 1.2.) A further approximation to r is then obtained by
solving in x,
(2.3)
This yields the second approximation,
(2.4)
x - x _ f(xo)
I - 0 f'(x
o
) .
The process is then repeated at Xl leading to a new value x
2
, and so
on, as illustrated in Fig. 1.2. The general recurrence relation is
(2.5)
It is clear from Fig. 1.2 that
(2.6) Xo < Xl < X
2
< ... < r,
if Xo < r initially. Analytically, this follows from the inequalities
!(x
n
) > 0, f'(x
n
) < O. This monotonicity is an important property
computationally. We would prefer to have alternate approximations
above and below r, that is, Xo < X
2
< X
4
< ... < r < ... < Xs < Xl>
but this is usually very difficult to arrange.
A second property, even more important computationally, is
perhaps not so obvious. It is what is called quadratic convergence.
We assert that
( 2.7)
where k is independent of n, under a reasonable hypothesis con-
f'''(x).
10 QUASILINEARIZATION AND NONLINEAR BOUNDARY.VALUE PROBLEMS
To see this, write
(2.8)
f(xn)
Xn+l - r = Xn - f'(Xn) - r
f(xn) (' f(r)
= Xn - f'(Xn) - r - f'(r)
= qJ(Xn) - qJ(r),
where qJ(x) = x - j(x)Jj'(x). Using the first three terms of the
Taylor expansion with a remainder, we obtain
(2.9)
(x r)2
Xn+l - r = (xn - r)g/(r) + n - qJ"(O),
2
where Xo ~ Xn ~ 0 ~ r. Since
(2.10) cp'(x) =f(x)!"(x)
f'(X)2 '
we see that qJ'(r) = O.
Hence
(2,11)
where k = max"'o::::o::::r Icp"(0)IJ2, a bound depending uponj'l/(x).
Furthermore,
(2.12) x
n
+
1
- xn = qJ(xn) - cp(x
n
-
1
)
( )
'() (xn - X
n
_
1
)2 "(0)
= xn - x
n
-
1
cp x
n
- 1 + 2 qJ,
where X
n
-
1
~ 0 ~ x
n
, which leads to
Referring to (2.5), we replace j(xn-1)!f'(X
n
-
1
) by Xn - x
n
-
1
. Hence,
(2.13) yields
(2.14)
Thus,
(2,15)
where
(2.16)
kl = max [1f"(0) 1 + 1 qJ'f(O) I] .
x095,r 11'(0)1 2
The relation in (2.15) is also called quadratic convergence.
II
It follows that as Xn approaches r, there is an enormous accelera-
tion of convergence. Asymptotically, each additional step doubles
the number of correct digits in the approximation. This is a most
important property when large-scale operations are involved, not
only because the computing time is usually directly proportional to
the number of iterations, but also because round-off error can increase
in a serious fashion as the number of stages increases. It is interesting
then to conceive of computing as being analogous to summing a
divergent series which is asymptotic. There is an optimal partial sum
for which the error is minimum.
3. MULTIDIMENSIONAL VERSION
The method readily generalizes to higher dimensions. Consider
the problem of determining a solution of a system of simultaneous
equations
(3.1) i = 1,2, ... , N,
or, in vector notation,
(3.2)
J(x) = O.
If Xo is an initial approximation, write
(3.3)
where ](xo) is the Jacobian matrix
(3.4) J(x
o
) = (OJi) .
OX; "'="'0
The new approximation is thus
(3.5) Xl = Xo - J(xo)-Y(xo),
and, generally, we obtain the recurrence relation
( .l.6)
It is now not difficult to obtain the analogues of the results of the
previous section under appropriate assumptions concerningf(x) and
its partial derivatives.
Analogous results hold for functional equations, but require more
sophisticated concepts both to state and establish.
4. SQUARE ROOTS
As a particularly interesting application of the procedure discussed
in Sec. 2, consider the problem of determining the positive solution
of x2 = y, y > 0, which is to say, that of obtaining the square root
ofy.
Settingf(x) = x
2
- y, the recurrence relation becomes
(4.1)
x = x _ (x; - y) = Xn + L
n+l n 2 2 2 '
Xn xn
a very well-known algorithm, ascribed to Heron of Alexandria.
The demonstration of monotonicity and quadratic convergence
by no means disposes of all problems connected with the use of this
algorithm. Suppose that we are required to extract the square roots
of a set of positive numbers Yb Y2, ... , YN, lying in a fixed interval
[l,a]. Suppose that we don't know what these numbers will be in
advance, but that we are required to choose an initial approximation
Xo in advance. What is the best fixed initial approximation to choose?
One way to formulate this question precisely is to agree that we
will iterate (4.1) exactly N times and ask for the value of Xo that
minimizes the function
(4.2)
Let xo(N) denote a value (actually the value) that minimizes. It can
be shown that
(4.3) lim xo(N) = al-l.
N"'<:IJ
This result is due to Sylvester and Hammer. Analogous results can be
derived for the algorithm determined by the problem of extracting a
kth root, k > 1, namely
(4.4)
The corresponding limit for xo(N) can be readily guessed, but it has
not been established rigorously. The analogous problems for more
general algebraic, transcendental, and functional equations have
apparently never been discussed, or even posed.
5. QUASILINEARIZATION
Let us now consider the equation
(5.1) x + bxn = a, a, b > 0, n> 1,
and indicate how quasi linearization techniques permit us to obtain
a representation for the unique positive solution in terms of the
maximum operation.
We begin with the observation that
(5.2)
for x ;;:: 0 and n > 1, with the maximum attained at y = x. Subse-
quently, we shall examine the source of this result. At the moment,
let us accept it as an easily established identity.
Then (5.1) may be written
( 5.3) x + max [bnxyn-l - ben - l)yn] = a,
v2:0
since b > 0, or, equivalently,
( 5.4)
x + bnxyn-l - ben - l)yn S a
for all y ~ O. Hence, since 1 + bnyn-l ;;:: 0,
( 5.5)
x S a + ben - 1)yn
1 + bnyn-l
for all y ~ o. Since equality is attained for y = x, we may write
(5.6)
. [a + ben - l)yn]
x = mln ,
'V 1 + bnyn-l
an explicit analytic representation for the desired solution.
Expressions of this type are useful since they furnish quick and
convenient upper bounds by appropriate choice of y.
6. THE RICCATI EQUATION
Let us now apply this technique of quasilinearization to an
important functional equation, the Riccati equation. The analysis
that follows not only has tutorial value, but is useful in its own right,
since, as we have mentioned above, the Riccati equation, together
with its multidimensional and function space analogues, plays a
fundamental role in modern control theory and mathematical
physics, particularly in connection with dynamic programming and
invariant imbedding.
The Riccati equation is a first-order nonlinear differential equation
having the form
(6.1) v' + v
2
+ p(t)v + q(t) = 0.
Despite its rather simple form, it cannot be solved explicitly in terms
of quadratures and the elementary functions of analysis for arbitrary
coefficients p(t) and q(t).
The connection with the second-order linear differential equation
is at first sight purely formal. Starting with the equation
(6.2)
we set
(6.3)
u" + p(t)u' + q(t)u = 0,
u'
v =_.
t-
and obtain (6.1) as the equation for v.
THE RICCATI EQUATION IS
As the reader will see subsequently in our discussion of variational
problems via dynamic programming, the connection between the
two equations is much more profound and meaningful, and more
varied.
7. THE FIRST-ORDER LINEAR EQUATION
Prior to our study of (6.1) via quasilinearization and successive
approximations, let us for the sake of completeness recall the solution
of the first-order linear equation
(7.1)
dw
- - f(t)w = get),
dt
w(O) = c.
Multiplying through by the function f(,,) dS, an integrating factor,
we obtain the equation
(7.2)
which we solve by immediate integration.
Hence, the solution of (6.1) is given by the expression
(7.3) w = cefJf(s)dS + ds,
a result we shall use below.
In place of this rather cumbersome formula, let us write
(7.4) w = T(j,g) ,
which defines the operation T(f,g), a linear inhomogeneous operation
on g. Observe that the positivity of the exponential function allows
us to conclude that gl(t) g2(t) for t 0 implies T(f,gl) T(f,g2)
for t 0, a most important and useful property, which provides
the basis for the subsequent results of this chapter.
8. SOLUTION OF RICCATI EQUATION IN TERMS
OF MAXIMUM OPERATION
Let us now show that the Riccati equation can be solved in terms
of a maximum operation. Consider the parabola y = X2 and the
Y - Y, = 2 X, (x - x,)
Y
Y = 2XI X - x,'
X
Fig. 1.3
(8.1)
or
(8.2)
(See Fig. 1.3.)
Since the curve y = X2, as a consequence of its convexity, lies
above the tangent line, we have the inequality
(8.3)
for all Xl, with equality only if X = Xl' Hence,
(8.4)
a result which can, of course, be established immediately in many
ways. This simple geometrical idea is at the bottom of the results of
(5.2) and (4.4).
Returning to the equation of (6.1),
(8.S) v' - _v
2
- p{t)v - q(t),
let us replace the quantity v
2
by its equivalent expression,
max [2uv - u
2
],
u
so that the equation now has the form
(8.6) v' = -max [2uv - u
2
] - p(t)v - q(t)
u
= min [u
2
- 2uv - p(t)v - q(t)].
u
17
Were it not for the minimization operation, the equation would be
linear. As it is, the equation possesses certain properties ordinarily
associated with linear equations. It is for these reasons that we employ
the term "quasilinearization."
Consider the companion equation, the linear differential equation
(8.7) w' = u
2
- 2uw - p(t)w - q(t),
where u(t) is now a fixed function of I. Let v and w have the same
initial values, v(O) = w(O) = c.
Let vet), the solution of the Riccati equation, exist in the interval
[0,10]. A proviso concerning existence is necessary, since the solution
of a Riccati equation need not exist for all I. Consider, for example,
the equation u' = 1 + u
2
, u(O) = 0, with the solution u = tan t
for 0 S 1 < 1T/2.
To establish the inequality w ~ v, we argue as before. Observe
that (8.6) is equivalent to the equation
(8.8)
for all u(I), or
( X.9)
v' S u
2
- 2uv - p(t)v - q(t)
v' = u
2
- 2uv - p(t)v - q(t) - ret),
where r(t) ~ 0 for 1 ~ O. The function r(t) depends, of course,
upon u and v, but this is of no consequence at the moment.
Hence, using the notation of Sec. 7,
( X.tO)
v = T( -2u - pet), u
2
- q(t) - r e t ~
~ T( -2u - pet), u
2
- q(t = w,
using the positivity property of the operator T(f,g).
Since this inequality holds for all u(t), with equality for u(t) = vet),
we have the desired result:
(8.11) vet) = min T( -2u - pet), u
2
- q(t,
u
which we state as a theorem:
Theorem 1. For 0 S t S to, where to depends upon the interval of
existence of vet), we have the representation
(8.12) vet) =
+ fe-f!(2U(rl+P(rdr[u2(s) - q(s)] dS]'
Using quasilinearization, we have obtained an analytic expression
for the solution of the Riccati equation.
9. UPPER AND LOWER BOUNDS
From the representation of (8.12), we can readily obtain upper
bounds for v(t) by appropriate choice of u(t). From this, in turn, we
can deduce upper bounds for the solution of the second-order linear
(9.1)
w" + p(t)w
'
+ q(t)w = 0,
recalling the discussion of Sec. 7.
If q(t) < 0, we can obtain lower bounds by replacing v = Ijw and
proceeding as above with the equation
(9.2) Wi - 1 - p(t)w - q(t)w
2
= 0.
These techniques have been used to study the asymptotic behavior
of the solution as a function of a parameter and as a function of t,
notably by Calogero.
Furthermore, to obtain more precise estimates, we can use more
general transformations of the form
(9.3)
since w will satisfy a Riccati equation whenever v does.
10. SUCCESSIVE APPROXIMATIONS VIA
QUASILINEARIZATION
19
We are now in a position to use the formula in Theorem 1 of Sec. 8
to generate a series of approximations to the solution v(t). We know
that the minimum is attained by the solution v(t) itself. Hence, we
suspect that if Vo is a reasonable initial approximation, then VI ob-
tained as the solution of
(10.1) v{ = v ~ - 2VOVI - p(t)v
1
- q(t),
will be even a better approximation, by analogy with the procedure
we employed to find the root of I(x) = 0, using the Newton-Raphson
approximation. Repeating this approach, we obtain the recurrence
relation
This is precisely the recurrence relation we would obtain if we applied
a Newton-Raphson-Kantorovich approximation scheme to the
nonlinear differential equation
(10.3) v' = _v
2
- p(t)v - q(t), v(O) = c.
Generally, if the differential equation were v' = g(v,t), we would use
the approximation scheme
We wish to emphasize, however, that the two approaches are not
equivalent despite the fact that they coincide in particular cases. As
we have indicated in the Introduction and shall discuss in some detail
in Chapter Seven, quasi linearization owes a great deal to its inception
within the theory of dynamic programming, with the concept of
approximation in policy space playing an important and guiding role.
In particular, from this viewpoint we are led to expect monotonicity
of convergence. On the other hand, the geometric background of the
Newton-Raphson approximation procedure leads us to expect quad-
ratic convergence. As we shall demonstrate below, both properties
20 QUASI LINEARIZATION AND NONLINEAR BOUNDARYVALUE PROBLEMS
are valid, and the resultant combination is extremely powerful
computationally and analytically.
Alternatively, we can regard quasilinearization as a systematic
exploitation of duality.
II. MONOTONICITY
Let us now establish the monotonicity of approximation. Consider
the relations
(11.1) v ~ + 1 = v; - 2VnVn+1 - p(t)v
n
+1 - q(t), V
n
+1(O) = c,
~ V ~ + I - 2vn+1vn+1 - p(t)Vn+1 - q(t)
(since the minimum value of the right-hand side is achieved for
Vn = v
n
+
I
). We deliberately do not simplify in order to emphasize
the relation to the equation
Comparing this equation with the inequality of (10.1), and using the
property of the linear operator discussed in Sec. 7, we see that
(11.3)
Hence, inductively,
(11.4)
If the function Vo is chosen arbitrarily, it mayor may not be true
that Vo ~ VI.
To show that vn(t) ~ v(t) for all n ~ 1, we write
(11.5) Vi = _v
2
- p(t)v - q(t)
= min [w
2
- 2wv - p(t)v - q(t)]
w
~ v ~ - 2vnv - p(t)v - q(t).
Hence, once again using the property of the linear operator, v ~ v"
for each n ~ 1.
Consequently, in any interval [O,t
o
] within which v(t) exists, we
can assert the convergence of the sequence {v,,}.
From the relation
(11.6) v,,+1(t) = -2VnVn+1 - p(t
1
)V
n
+
1
- q(t
1
)] dt
1
,
we conclude, on the basis of bounded convergence, that
(11.7) vet) - c = f[ _v
2
- P(t1)V - q(t
1
)] dt
1
,
where v = 00 vn(t). Hence vet), as an integral, is continuous.
Differentia ting,
(11.8) v' = _v
2
- p(t)v - q(t), v(O) = c.
A standard uniqueness theorem shows that v = v.
Finally, since the elements of the sequence {v
n
} are continuous,
since the convergence is monotone, and since v(t) is continuous,
we can assert, using Dini's theorem, that the convergence is actually
uniform in [O,t
o
].
12. DISCUSSION
In the previous sections, we have discussed two important prop-
erties of the sequence of approximations obtained by means of quasi-
linearization, quadratic convergence and monotonicity. Of these, the
first was an immediate consequence of the generalized tangent approx-
imation, a Newton-Raphson-type procedure. The second, however,
holds only for special classes of operators or for small t-intervals. It
is a consequence, in this case, of the positivity property of the operator
(12.1)
d .
L= dt + p(t),
or, more properly, of the operator L-1.
We have repeatedly used the fact that
(12.2) Lu O=>u O.
Fortunately, as we shall observe in what follows, many of the most
important operators of analysis and mathematical physics possess
this positivity property. In the following pages we will treat second
and nth-order linear differential equations, parabolic partial differen-
tial equations, elliptic partial differential equations, and some impor-
tant types of quasilinear partial differential equations.
Extensions of the Newton-Raphson technique to functional space
have been made by Kantorovich and others. In view of his contribu-
tions to this important area, it is fitting to call the approximation
technique used above an application of a general Newton-Raphson-
Kantorovich technique.
13. A FUNDAMENTAL LEMMA
The positivity property of L -1 can be used to establish the following
fundamental result in the theory of differential equations.
Theorem 2. If u(t), vet) ~ for t 2 0, c ~ 0, then
(13.1)
implies that
(13.2)
u(t) ~ c + fU(s)V(S) ds
This result, together with extensions and generalizations, is
extremely useful in establishing existence and uniqueness of solutions
of nonlinear differential equations and in studying the stability of
solutions of ordinary and partial differential equations. It belongs to
the general area of differential inequalities, a theory inaugurated by
Caplygin.
Chapter One
COMMENTS AND BIBLIOGRAPHY
2. For a detailed discussion of the historical development of the Newton-
Raphson method, see
D. T. WHITESIDE, "Patterns of Mathematical Thought in the Latter Seventeenth
Century," Arch. Hist. of the Exact Sci., Vol. I, No.3, 1961, pp. 179-388
(especially p. 207).
23
For analytic results and extensions to function space, see
L. V. KANTOROYICH, "Functional Analysis and Applied Mathematics," Uspehi
Mat. Nauk, Vol. 3, 1948, pp. 89-185,
L. V. KANTOROYICH and V. I. KRYLOY, Approximate Methods of Higher Analysis,
Interscience Publishers Inc., New York, 1958,
M. M. V AINBERG, Variational Methods for the Study of Nonlinear Operators,
Holden-Day, Inc., San Francisco, 1964,
S. S. SAYENKO, "Iteration Method for Solving Algebraic and Transcendental
Equations," Z. Vycisl. Mat. Mat. Hz., Vol. 4, No.4, 1964, pp. 738-744,
where many further references will be found.
3. The min-max problem posed in this section was discussed in
J. J. SYLVESTER, "Meditations on Idea of Poncelet's Theorem," and "Notes on
Meditation ... ," Math. Papers, Vol. II,
and in
P. C. HAMMER, Proceedings Compo Seminar, IBM, December 1949, p. 132.
As far as we know, these are the only published works in this area, although
unpublished notes by Bellman and Tukey exist.
5. This result was given in
R. BELLMAN, "On Explicit Solutions of Some Trinomial Equations in Terms of
the Maximun Operation," Math. Mag., September-October 1956, pp. 41-44.
6. The theory of the solubility of differential equations in terms of the ele-
mentary functions and quadratures was created by Liouville and further devel-
oped by Ostrowski and Ritt. For an exposition, see
J. F. RITT, Integration in Finite Terms, Liouville's Theory of Elementary Methods,
Columbia University Press, New York, 1948.
G. N. WATSON, Bessel Functions, Cambridge University Press, 1944, pp. 111-123.
See also
G. H. HARDY, The Integration of Functions of a Single Variable, Cambridge
University Press, 1958.
8. This result was given in
R. BELLMAN, "Functional Equations in the Theory of Dynamic Programming-V:
Positivity and Quasilinearity," Proc. Nat. A cad. Sci. USA, Vol. 41, 1955,
pp. 743-746.
It was treated in considerable detail in
R. KALABA, "On Nonlinear Differential Equations, the Maximum Operation,
lind Monotone Convergence," J. Math. Mech., Vol. 8, 1959, pp. 519-574.
The explicit representation of the solution of the Riccati equation has been
used by Calogero to attack some important questions in scattering theory;
see
F. CALOGERO, "A Novel Approach to Elementary Scattering Theory," Nuovo
Cimento, Vol. 27, 1963, pp. 261-302.
---, "A Variational Principle for Scattering Phase Shifts," Nuovo Cimento,
Vol. 27, 1963, pp. 947-951.
--, "A Note on the Riccati Equation," J. Math. Phys., Vol. 4, 1963, pp.
427-430.
---, "The Scattering of a Dirac Particle on a Central Scalar Potential,"
Nuovo Cimento, Vol. 27, 1963, pp. 1007-1016.
---, "Maximum and Minimum Principle in Potential Scattering," Nuovo
Cimento, Vol. 28, 1963, pp. 320--333.
For an application to stability theory, see
R. BELLMAN, "A Note on Asymptotic Behavior of Differential Equations,"
Boll. d'Unione Mate., Vol. 18, 1963, pp. 16-18.
10--11. These results were first given in the work by R. Kalaba mentioned
above. For an application to stochastic differential equations, see
R. BELLMAN, "On the Representation of the Solution of a Class of Stochastic
Differential Equations," Proc. Amer. Math. Soc., Vol. 9, 1958, pp. 326-327.
For a generalization of the preceding results to the equation u' = g(u,t),
see
R. BELLMAN, "On Monotone Convergence to Solutions of u' = g(u,t)," Proc.
Amer. Math. Soc., Vol. 8, 1957, pp. 1007-1009.
For a very interesting new use of second-order approximation techniques, see
J. MOSER, "A New Technique for the Construction of Solutions of Nonlinear
Differential Equations," Proc. Nat. Acad. Sci. USA, Vol. 47, 1961, pp. 1824-
1831.
13. This result was first used for the study of stability in
R. BELLMAN, "The Boundedness of Solutions of Linear Differential Equations,"
Duke Math. J., Vol. 14, 1947, pp. 83-97.
---, Stability Theory of Differential Equations, McGraw-Hill Book Company,
Inc., New York, 1953.
Extensions of this lemma are due to Bihari and Langenhop; see the
references to them and to Caplygin in
E. F. BECKENBACH and R. BELLMAN. Inequalities. Springer, Berlin. 1961.
For applications of this lemma to the study of existence and uniqueness theorems
for ordinary differential equations. see
V. V. NEMYTSKI and V. V. STEPANOV, Qualitative Theory of D(tferential Equations,
Moscow, 1949 (in Russian).
For general discussions of nonlinear problems, see
T. VON KARMAN, "The Engineer Grapples with Nonlinear Problems," Bull.
G. TEMPLE, "Linearization and Delinearization," Proc. International Congress
of Math., 1958, pp. 233-247.
For some applications, the reader may consult
J. R. RADBILL, "Application of Quasilinearization to Boundary-Layer Equa-
tions," AIAA Journal, Vol. 2, No. 10, 1964, pp. 8-10.
K. S. P. KUMAR and R. SRIDHAR, "On the Identification of Control Systems
by the Quasilinearization Method," IEEE Trans. on Automatic Control,
Vol. AC-9, April 1964, pp. 151-154.
M. C. SMITH, C. S. Wu, and H. S. SCHWIMMER, Magnetohydrodynamic-
Hypersonic Viscous and Inviscid Flow Near the Stagnation Point ofa Blunt
Body, The RAND Corporation, RM-4376-PR, January 1965.
R. KALABA, "Dynamic Programming, Fermat's Principle, and the Eikonal
Equation," J. Optical Soc. Amer., Vol. 51, October 1961, pp. 1150-115\.
Chapter Two
TWOPOINT BOUNDARYVALUE PROBLEMS
FOR SECONDORDER DIFFERENTIAL
EQUATIONS
I. INTRODUCTION
We now turn our attention to the study of nonlinear second-order
differential equations of the form
(1.1) un = g(u,u',t)
with the two-point boundary condition u(O) = ah u(b) = a
2
We
possess no convenient or useful technique for representing the general
solution in terms of a finite set of particular solutions as in the linear
case (a point we shall discuss below). Consequently, we possess no
ready means of reducing the transcendental problem involved in
solving (1.1) to an algebraic problem as is the situation in the case
where g(u,u' ,f) is linear in u and u'. Indeed, the entire area of existence
and uniqueness of solutions of nonlinear differential equations is one
of great complexity and subtlety, and one which contains many
uncharted regions.
To obtain an analytic foothold, and simultaneously to provide
computational algorithms, we must have recourse to approximation
techniques. Fixed-point methods, so valuable in establishing existence
of solutions, are of no use numerically. Generally, few of the standard
classical techniques of successive approximations are of much utility
numerically. Nonetheless, as of now, a number of powerful compu-
tational methods exist. We shall, however, concentrate upon one
particular method-that furnished by the theory of quasilinearization.
Our aim is to obtain the solution of (1.1), when it exists, as the
limit of a sequence of solutions of linear differential equations. Each
of these can be solved numerically in a convenient fashion using the
27
28 QUASI LINEARIZATION AND NONLINEAR BOUNDARY-VALUE PROBLEMS
fundamental tool of superposition. Furthermore, as in the case of the
Riccati equation discussed in the previous chapter, we want quadratic
convergence, and, if possible, monotonicity. The results we obtain
will show that these are not unreasonable demands.
2. TWOPOINT BOUNDARYVALUE PROBLEMS FOR
SECONDORDER LINEAR DIFFERENTIAL EQUATION
Preparatory to a study of general second-order differential equa-
tions, let us examine the linear equation
(2.1) u" + p(t)u' + q(t)u = 0,
u(o) = aI'
u(b) = a
2
,
a two-point boundary-value problem. We have considered specific
simple boundary conditions in place of general conditions at the
end-points, since the approach is the same in all cases.
Let U
I
and U
2
be the two principal solutions of the homogeneous
equation above, defined by the initial conditions
(2.2) ul(o) = 1,
u{(o) = 0,
u
2
(0) = 0,
u ~ ( o ) = 1.
Then, by virtue of the linearity of (2.1), every solution can be repre-
sented as a linear combination
(2.3)
where CI and C
2
are constants determined by the initial or boundary
conditions. Using the conditions at t = 0 and t = b, we readily obtain
the two linear algebraic equations
(2.4)
Hence,
(2.5)
TWOPOINT BOUNDARYVALUE PROBLEMS 29
Thus, if u
2
(b) =;tf 0, there is a unique solution. If u2(b) = 0, there is a
solution only if a
2
= aiui(b), and in this case there is a one-parameter
family of solutions of the form
(2.6)
with C2 arbitrary. Hence, existence and uniqueness are here strongly
interconnected, as is frequently the case.
T ~ e condition u
2
(b) =;tf 0 we recognize as a characteristic value
condition associated with the Sturm-Liouville equation
(2.7) u" + p(t)u' + Aq(t)U = 0,
u(o) = u(b) = 0.
In our case, we demand that A = 1 not be a characteristic value.
In other words, we insist that when A = 1 the only solution of (2.7)
be the trivial solution u(t) = O.
3. THE INHOMOGENEOUS EQUATION
Let us now turn to the inhomogeneous equation
(3.1) u" + p(t)u' + q(t)u = ret),
. u(o) = ai'
Taking advantage of the linearity, we write
(3.2)
u = v + w,
where wand v are chosen, respectively, to satisfy the equations
(3.3)
w" + p(t)w' + q(t)w = ret),
w(o) = 0, w'(O) = 0,
and
( 3.4)
v" + p(t)v' + q(t)v = 0,
v(b) = a
2
- web).
The first equation is an initial value problem which we consider below,
and the solution of the second has already been discussed.
It remains to discuss the analytic solution of the inhomogeneous
equation of (3.3) in terms of the solutions of the homogeneous equa-
tion of (2.1). The most elegant and efficient way of doing this is to
use vector-matrix theory, a method we shall sketch in the following
section.
Let us, however, to avoid detours, employ the ingenious variation
of parameters technique of Lagrange. Let Ul and U2 be particular
solutions of the homogeneous equation, as above, and write
(3.5)
where Sl and S2 are functions of t to be determined at our convenience.
Then
(3.6)
To simplify, set
(3.7)
This is the first condition on and Since w' = + with
this constraint, we have
(3.8)
Thus
(3.9) w" + p(t)w' + q(t)w = + + q(t)u
1
)
+ + + q(t)u
2
)
+ + = ret).
Since U
1
and U2 satisfy the homogeneous equation, this reduces to
(3.10)
+ = ret).
Combining this with (3.7), we have two simultaneous linear algebraic
equations for and A solution exists, provided that the deter-
minant (the Wronskian)
(3.11)
TWO-POINT BOUNDARY-VALUE PROBLEMS
31
is nonzero_ Accepting this fact, which we will establish in Chapter
Three, Sec_ 3, we have
(3.l2)
I
0 u21
, ret)
S -
1 - Wet) ,
l
UI 0 I
, u{ ret)
S2 == .
Wet)
Hence,
(3.13)
Sl = _ (t u2(t
l
)r(tl ) dtl ,
Jo W(t
l
)
S2 == (t ul ( tl)r( tl ) dt l ,
Jo W(t
l
)
choosing Sl(O) == S2(0) == 0, since we want to satisfy the conditions
w(O) == w'(O) == O. '
Thus,
(3.l4)
1
t ( )[ L U (...t
W == r tl -
o W(t
l
)
= f G(t,tl)r(t
l
) dt
l

To solve the equation
(3.15) U" + p(t)u' + q(t)u == ret), u(O) == u(b) == 0,
we proceed as follows. The general solution of (3.15) has the form
(3.16)
where C
2
is a constant to be determined. Setting t == b, we see that
(3.l7) C2 == _ (b G(b,tl)r(tl) dt l .
Jo u
2
(b)
Hence,
(3. t 8)
A simple calculation shows that we may write
(3.19) u(t) = Lb K(t,t1,b)r(t
1
) dtl>
where
(3.20)
Conversely, differentiation shows that if u satisfies (3.19), it satisfies
(3.16).
4. VECTORMATRIX APPROACH
Let us point out in passing that a straightforward derivation of
(3.14) may be obtained using a minimum of vector-matrix theory
applied to linear differential equations.
We begin with the observation that
(4.1) u" + p(t)u' + q(t)u = ret)
may be written in the form of a first-order linear system
(4.2)
u ~ = - p(t)u
2
- q(t)u
1
+ ret).
Hence, introducing respectively the vectors u(t) and !(t), and the
matrix A(t) defined by
(4.3) x(t) = [u1(t)], J(t) = [0 ],
u
2
(t) ret)
A(t) = [0 1]
-q(t) - pet) ,
we have, equivalent to (4.2), the vector-matrix system
(4.4) x' = A(t)x + J(t),
with an initial condition x(O) = O.
There are two ways of approaching the solution of (4.4). The
first uses the Lagrange variation of parameters approach in vector-
matrix form. The second, which generalizes to any type of linear
functional equation, uses the solution of the adjoint equation. Let us
discuss the second approach.
Let yet) be an as yet unspecified matrix and start with the relation
(4.5) LT Yx' dt = LT [Y A(t)x + Yf(t)] dt.
Integrating by parts, we have
(4.6) Y(T)x(T) - Y(O)x(O) - LT Y'x dt = LT Y A(t)x dt + LT Yf(t) dt.
Since x(o) = 0, this may be written
(4.7) Y(T)x(T) = LT[y, + Y A(t)]x dt + JoT Yf(t) dt.
Since this holds for any matrix Y, let us choose a most convenient
one. We specify the matrix Y(t) = Y(t,T), in the interval [O,T], by
the equation
(4.8) Y' + Y A(t) = 0, YeT) = I.
With Y(t) determined in this fashion, we have
(4.9) x(T) = LT Yf(t) dt.
If X(t) denotes the solution of the matrix equation
( 4.10) X' = A(t)X, X(O) = I,
it is easy to see that the solution of (4.8) is given by
( 4.11)
This is a consequence of the relation
( 4.12)
.!!. (X-I) = _ X-I dX X-I.
dt dt
Hence, (4.9) becomes
(4.l3) x(T) = LT X(T)X(trlf(t) dt.
Carrying out the indicated matrix calculations, we obtain the explicit
result of (3.14). The preceding expression, in (4.13), is preferable for
analytic purposes, since the structure of the representation is quite
clear.
5. GREEN'S FUNCTIONS
The function K(t,t1b) introduced in Sec. 3 is a very important
function, called a Green's function. It is determined both by the form
of the particular linear equation and by the boundary conditions.
For example, if the equation is
(5.1) u" = ret), u(o) = u(b) = 0,
a simple calculation along the preceding lines shows that
(5.2) u = Lb K(t,tI,b)r(tI) dtl>
where
(5.3) K(t,tl,b) = t(t
i
- b)/b, I:::;; 11 :::;; b,
= t
I
( t - b )/b, 0:::;; tl :::;; I.
We shall employ this result below.
There are many interesting and important properties of Green's
functions which we will not be able to discuss here. They playa basic
role in the advanced theory of differential equations and their asso-
ciated integral equations. We shall, however, in the following chapter
discuss some important questions of positivity of Green's functions_
6. CONVEXITY
Let us now generalize the representation
(6.1) u
2
= max (2uv - v
2
),
v
TWO-POINT BOUNDARY-VALUE PROBLEMS 35
y = f (x)
--------------4--------------------- x
Fig. 2.1
used in Chapter One to study the Riccati equation, by introducing
the concept of convexity. We say that a function j(x) is convex for
a ~ x ~ b ifj"(x) ~ 0 in this interval. It is strictly convex ifj"(x) > o.
Consider the graph of a strictly convex function (see Fig. 2.1),
and consider the tangent at the point P(u,j(u)),
(6.2) y - feu) = f'(u)(x - u).
By virtue of strict convexity, we see that the curve is always above
the tangent for all values of x except x = u. Hence,
(6.3) f(x) ~ feu) + f'(u)(x - u),
for all u and x with equality at x = u.
Therefore, we may write
( 6.4) f(x) = max [feu) + (x - u)f'(u)].
u
The result can, of course, be obtained purely analytically, by the use
of calculus.
Similarly, for a function of two variables, we have
(6.5) f(x,y) = max [f(u,v) + (x - u)fu + (y - v)fv],
U.l'
provided that j(x,y) is strictly convex in the domain of interest,
which is to say, provided that
(6.6)
fxx > 0,
I
fxx jXY I > 0,
fxy fyy
or equivalently, that the quadratic form
(6.7)
is positive definite.
Analogous results hold for functions of any finite number of vari-
ables and for functionals. The general discussion of convexity re-
quires some subtlety but merits it, since applications of this simple,
yet fundamental, concept occur throughout all of analysis and
geometry.
We are now ready to employ quasilinearization for analytic and
computational purposes. We shall first use this technique to obtain
quadratic convergence, and illustrate this by means of some examples.
Let us begin, for the sake of illustration, with the equation
(7.1) u" = feu), u(o) = u(b) = 0.
Let uo(x) be some initial approximation and consider the sequence
{un} determined by the recurrence relation
(7.2)
Each function un(x) is a solution of a linear equation, a very important
characteristic of this algorithm.
This is an application of the Newton-Raphson-Kantorovich
approximation method in function space. On the basis of what has
gone before, we can expect quadratic convergence, where. convergence
occurs.
8. DISCUSSION
Since each solution is determined by a two-point boundary con-
dition, it is not evident a priori that the sequence {Un} as defined by
(7.2) actually exists.
We shall first demonstrate that the sequence is well defined,
provided that we restrict x to a sufficiently small interval [O,b], and
then that the sequence converges quadratically. In the following
chapter, we shall consider the proof of monotone convergence,
which requires some more sophisticated ideas, and use this result to
enlarge the interval of convergence of the sequence of approxima-
tions. We shall present a variety of approaches to this very important
and interesting problem area of monotonicity and positivity.
9. EXISTENCE AND BOUNDEDNESS
To begin with, we shall establish the existence and uniform
boundedness of the sequence {un(x)} for b sufficiently small. Starting
with the equation of (7.1), namely
(9.1)
U ~ + l = f(u
n
) + (un+! - un)f'(u
n
),
un+1(0) = un+1(b) = 0,
we obtain, as illustrated in Secs. 4 and 5, the linear integral equation
(9.2) u
n
+
1
= i
b
K(x,y)fJ(u
n
) + (u
n
+
1
- un)!'(u
n
)] dy,
where K(x,y) is the Green's function
(9.3) K(x,y) = x(y - b)/b,
= (x - b)y/b,
o ~ x < y ~ b,
b ~ x > y ~ o.
To obtain (9.2), we regard (9.1) as an equation of the form u" =
r(t), u(O) = u(b) = 0, and apply the result stated in Sec. 5.
Observe that
(9.4)
b
max IK(x,y)1 = -,
"'.11 4
where the maximization is over the region ~ x, y ~ b. Let
(9.5) max max (If(u)l, If'(u)I) = m,
lui :0;1
assuming that m < 00, and choose uo(x) so that I uo(x) I ~ I, for
~ x ~ b.
Turning to (9.2), we have
(9.6) lu
n
+
1
1 ~ f'K(X,Y)1 [If(un)1 + lunllf'(un)1 + If'(u
n
)llu
n
+1l] dy.
Hence, writing mi = maxO:O;X:O;b lu
1
(x)l, we have, for n = 0,
Provided, therefore, that b
2
mj4 < I, we obtain the bound
(9.8)
< b
2
m/2
m
1
- 1 _ b
2
m/4
This upper bound is itself less than 1 if b
2
~ 4j3m, a constraint that
can be met by taking the interval [O,b] to be sufficiently small,
for example, b
2
~ 4j3m. Thus, under these conditions, m
1
~ l.
It is clear that this procedure can be continued inductively with
the result that we can assert that lun(x)1 ~ 1 for ~ x ~ b, provided
that b
2
~ 4j3m. We have thus demonstrated that the inductive
definition of the sequence {uix)} is meaningful.
10. CONVERGENCE
The convergence proof follows standard lines, with the difference
that the convergence will turn out to be quadratic.
Returning to the recurrence relation of (9.1), let us subtract the
nth equation from the (n + 1)st,
(to.l) (Untl - un)" =f(u
n
} - f(u
n
-
I
) - (Un - Un-tV'(u
n
-
l
)
+ !'(Un)(Unti - Un)
Regarding this as a differential equation for U
n
+
1
- Un' and converting
into an integral equation as before, we have
(10.2) (un+! - Un) = f K(x,y){J(u
n
) - f(Un-l)
- (Un - Un-1)f'(U
n
-
1
)
+ !'(Un)(U
n
+! - Un)} dy.
The mean-value theorem tells us that
where () lies between U
n
-
1
and Un. Hence, letting
(10.4) k = max 1f"(u)l,
lui :0;1
we have, very much as before,
b (b[k 2 ]
(10.5) IU n+1 - unl ::::;; "4 Jo "2 (Un - Un-I) + m IU
n
+
1
- unl dy.
Hence, using the previous argument (9.7) and (9.8),
b
2
k/8 2
max Iu
n
+! - unl ::::;; 2 (max IU
n
- Un-I/) .
x 1 - b m/4 x
(10.6)
This shows that there is quadratic convergence if there is conver-
gence at all.
If the elements of a sequence {un}, n = 0, I, 2, ... ,satisfy the
relation
(10.7) n = 1,2, ... ,
a simple induction shows that
(10.8)
The convergence will depend therefore upon the quantity
b
2
k/8
kvo = 2 max lUI - uol,
1 - b m/4 x
( 10.9)
which by the standard procedures used above can be shown to be less
than one for b sufficiently small.
It is interesting to note that the proof shows that it is sufficient for
convergence that max", IU
n
+
1
- unl be small enough for some n.
Consequently, even if the interval [O,b] appears to be too large ini-
tially, we can always hope that a judicious initial approximation
uo(x) will keep lu1(x) - uo(x)1 sufficiently small. Here, physical and
engineering experience and intuition can be of great service.
If, then, the quantity in (10.7) is less than one, we have uniform
convergence of the sequence {un(x)} to a function u(x) satisfying the
equation
(10.10)
u = f K(x,y)f(u) dy.
From this we deduce that u satisfies the original differential equation.
Furthermore, the foregoing techniques show that
(10.11)
max lu - unl ~ k2 max lu - u
n
_
1
1
2
,
'" '"
from which uniqueness follows in the usual fashion.
We have gone into careful detail as to the existence and conver-
gence of the sequence of approximations in order to indicate what
the necessary steps are. Henceforth, in treating the more complex
functional equations of the later chapters, we shall omit the proofs of
existence and convergence. These require no additional concepts or
techniques, apart from some small knowledge of vector-matrix or
function space notation. Notice that the convexity assumption, of
importance in the representation theorem, plays no role in the
quadratic convergence.
II. CONVERGENCE OF THE PICARD ALGORITHM
The usual approach based upon the recurrence relation
(11.1 )
p':, I 1 = f( V
7l
),
41
also yields a convergent sequence with very much the same restric-
tions on f and b. The convergence, however, will in general only be
geometric, which is to say
(11.2)
where k < 1, or IV
n
- vn-II ~ knc
o
.
The advantage of quadratic convergence, of course, lies in rapidity
of convergence. We shall, moreover, show subsequently, in Chapter
Three, that quasilinearization also provides the maximum interval
of convergence of successive approximations in a number of impor-
tant cases.
12. A NUMERICAL EXAMPLE
The equation
(12.1)
is of current interest in connection with magneto hydrodynamics, and
its numerical solution will be considered in later pages using quasi-
linearization. Let us here discuss the one-dimensional version
(12.2) u" = e
U
, u(O) = u(b) = 0,
which possesses the considerable merit of having a unique and explicit
analytic solution. For b = 1, this has the form
(12.3) u(x) = -In 2 + 2In [c sec {c(x - t)/2}],
where c is the root of
(12.4)
~ 2 = c sec ( ~ )
which lies between 0 and TT/2, namely c = 1.3360557, to eight figures.
Let us recall that we obtain (12.2) by means of multiplication by
u
'
and integration of both sides. This procedure yields u' in terms of
u, and a further integration achieves the desired explicit solution of
( 12.3).
Applying the techniques presented above, we consider the sequence
of approximations defined by
Taking a most obvious initial approximation uo(x) = 0, we compute
Table 2.1
THE APPROACH OF un(x) TO u(x)
x uo(x) u
1
(x) u
2
(x) u(x)
0.0 0.000000 0.000000 0.000000 0.000000
0.1 0.000000 -0.041285 -0.041436 -0.041436
0.2 0.000000 -0.072974 -0.0732683 -0.0732686
0.3 0.000000 -0.095386 -0.095800 -0.095800
0.4 0.000000 -0.108743 -0.109238 -0.109238
0.5 0.000000 -0.113181 -0.113704 -0.113704
0.6 0.000000 -0.108743 -0.109238 -0.109238
0.7 0.000000 -0.095386 -0.095800 -0.095800
0.8 0.000000 -0.072974 -0.0732683 -0.0732686
0.9 0.000000 -0.041285 -0.041436 -0.041436
1.0 0.000000 0.000000 0.000000 0.000000
the functions u
1
(x) and u
2
(x) as described above, using a Runge-
Kutta integration procedure.
Table 2.1 indicates the rapidity of convergence. The digits of
un(x) in agreement w i ~ h those of u(x) are italicized.
13. GENERAL SECONDORDER NONLINEAR
DIFFERENTIAL EQUATIONS
There is no difficulty in applying the same technique to the study
of the equation
(13.1) U" = !(u',u,x),
with nonlinear boundary conditions of the form
(13.2) gl(U(O),U'(O = 0, ga(u(b),u'(b = 0,
or, even more generally,
(13.3) gl(U(O),U'(O),U( b ),U'( b = 0,
g2(U(0),U'(0),u(b),u'(b = O.
43
We now apply quasilinearization to both the equation and the
boundary conditions. Thus, in the case of the equation of (13.1)
subject to the conditions (13.2), we generate the sequence {un(x)} by
means of the equation
(13.4)
with the linearized boundary conditions
(13.5) - un(O
+ - = 0,
and a similar equation derived from g2 for the point x = b.
There is no difficulty in establishing existence and uniqueness and
quadratic convergence for sufficiently small b, under obvious con-
ditions on the quantities ju, ju" giu, giu', i = 1, 2.
As a second example, consider the equation
(14.1) u(O) = u(b) = 0,
which arises in the study of the finite deflections of an elastic string
under a transverse load. * The explicit analytic solution for a
2
= 0.49
and b = 1 is
( 14.2)
u(x) = ..!. In [Cos a(x - 1
/
2)J.
a
2
cos al2
This equation was suggested to us by W. Prager.
Using the recurrence relation
(14.3) = 1 - +
U
n
+1(O) = u
n
+1(b) = 0,
with the initial approximation uo(x) = 0, we obtain the values shown
in Table 2.2.
Table I.I
THE CONVERGENCE OF un(x) TO u(x)
x uo(x) u
1
(x) u
2
(x) u(x)
0.0 0.000000 0.000000 0.000000 0.000000
0.1 0.000000 0.045000 0.046570 0.046571
0.2 0.000000 0.080000 0.082302 0.082304
0.3 0.000000 0.105000 0.107571 0.107573
0.4 0.000000 0.120000 0.122632 0.122635
0.5 0.000000 0.125000 0.127636 0.127639
0.6 0.000000 0.120000 0.122632 0.122635
0.7 0.000000 0.105000 0.107571 0.107573
0.8 0.000000 0.080000 0.082302 0.082304
0.9 0.000000 0.045000 0.046570 0.046571
1.0 0.000000 0.000000 0.000000 0.000000
15. CALCULUS OF VARIATIONS
A natural and most prolific source of nonlinear differential equa-
tions with two-point bpundary-value conditions is the calculus of
variations. Consider the problem of maximizing the functional
(15.1) J(u) = fg(U,U') dt,
subject to the initial condition u(O) = c. The vanishing of the first
variation produces the Euler equation
(15.2) og _ !!..( Og) - 0
ou dt ou'
with the terminal condition
(15.3) ag I = o.
au' /-b
Y
Po
(xo, Yo)
__ -_ P
,
(x" Yl)
X
o
45
Fig. 2.2-The path of a light ray through an inhomogeneous medium
If g is a quadratic form in u and u' plus a linear term in u, the Euler
equation is linear, and the two-point boundary condition can be
resolved in a straightforward fashion. In general, we face the usual
difficulties.
To illustrate the procedure of quasilinearization, consider the
following problem arising from an application of Fermat's principle.
Let the (x,y)-plane represent an optically inhomogeneous medium,
with the velocity of light at (x,y) denoted by v(x,y). A light particle,
initially at the point Po, passes through the point Pl' What path does
it take?
Pursuant to Fermat's dictum of minimum time, our task is to
minimize the integral
i
"'l ds 1"'1 .J 1 + y,2
(15.4) J(y) = -- = dx.
"'0 v(x,y) "'0 v(x,y)
(See Fig. 2.2.)
The Euler equation for this problem is
(15.5) :y [ -:X = o.
Let us take v(x,y) = y. This reduces the equation to
(15.6)
1 + y,2 + yy" = 0,
subject to the two-point boundary condition
(15.7)
It is known that the optimizing curves are arcs of circles with their
centers on the axis. We can use this to check the accuracy of our
numerical procedures.
Table 2.3
THE NUMERICAL RESULTS
x zo(x) zix) Z2(X) Z3(X)
1.00 1.00 1.000000 1.000000 1.000000
1.25 1.25 1.386195 1.391934 1.391941
1.50 1.50 1.651987 1.658305 1.658312
1.75 1.75 1.848866 1.854642 1.854050
2.00 2.00 2.000000 2.000000 2.000000
Write the in the form y" = -(1 + 1'2)/y, and take as an
initial approximation the function zo(x) obtained by drawing a
straight line joining Po and Pl. The general recurrence relation is
then
(15.8)
" (1 + + (1 + ( )
Zn+l = - 2 Zn+l - Zn
Zn
')
- - (Z n+l - Z n ,
Zn
with zn+1(x
o
) = Yo, Zn+l(X
l
) = Yl
Consider the particular problem where y(1) = 1, y(2) = 2. The
initial approximation is zo(x) = x. The numerical results obtained
in this way are shown in Table 2.3. (The program is given in
Appendix 1.)
The results for Z3(X) are in agreement with the exact solution to
six figures.
We can apply the quasi linearization technique in two ways. The
Euler equation in (15.2) is a nonlinear differential equation of a type
we have already encountered. Applying the customary linearization,
we obtain the linear equations
(16.1)
+ - un)
+ -
- !!:. + - Un)
dt
with the initial condition U
n
+1(O) = c and the linearized terminal
condition
(16.2) + (u
n
+
1
-
+ - = o.
Alternatively, we can proceed directly in the following manner.
Let Un be an approximation to the solution of the optimization
problem and consider the problem of minimizing the functional
(16.3)
I
n
= r h2(U,U
n
) dt,
where h
2
(u,u
n
) is the expansion of g(u,u',t) around the point (umu;,),
up to and including second-order terms. Thus,
(16.4) h2(U,U
n
) = + - un)
+ -
+ - U
n
)2
+ - un)(u' -
+ -
A straightforward calculation shows that the Euler equation
associated with I
n
is precisely that appearing in (16.1). We see then
that an approximate treatment of the exact variational equation is
equivalent, in this case, to an exact treatment of the approximate
variational problem. One advantage of the second approach lies in
the fact that dynamic programming and invariant imbedding tech-
niques can be applied to avoid two-point boundary-value problems
and the consequent necessity for solving linear systems of algebraic
equations. We shall discuss this point in detail below in Sec. 20 and
again in Chapter Seven.
17. UPPER AND LOWER BOUNDS VIA
QUASILINEARIZATION
Let us now briefly indicate how to apply quasilinearization to
obtain upper and lower bounds for certain classes of variational
problems. Consider, for illustrative purposes, the problem of mini-
mizing the functional
(17.1) J(u) = f(U'2 + f}J(u dt,
where u is ,subject to a condition such as u(O) = c. To simplify, let
us here take
(17.2) u'(l) = 0,
where the second condition is precisely the free boundary condition.
Assume that f}J(u) is convex so that we can write
(17.3)
Then
f}J(U) = max [f}J(v) + (u - v)f}J'(v)].
v
(17.4) min J(u) = min r\u,2 + f}J(u dt
u u Jo
= m ~ n m;x [f[U'2 + f}J(v) + (u - V)f}J'(V)]] dt.
It can be shown on the basis of quite general results in the theory of
games that the minimization and maximization operations can be
interchanged. Thus,
(17.5) min J(u) = maX min [r\u,2 + f}J(v) + (u - v)f}J'(v dt].
u v u Jo
The simplest proof of the validity of the interchange is a direct one,
based upon a direct calculation.
49
Po
Fig. 2.3
The whole point of the interchange is to obtain a lower bound for
leu), namely
(17.6) m:n J(u) ~ m:n [f(U'2 + q;(v) + (u - v)q;'(v dt]
for all functions vet). This minimum can now be determined ex-
plicitly, as we proceed to do below.
Before carrying out the calculation, let us indicate briefly the role
of duality, which we have been tacitly employing, in converting
minimization problems into maximization problems. Consider the
general problem of finding the minimum distance from a point po to a
convex set S (see Fig. 2.3). Consider the tangents to the boundary
curve B and the distance d from po to a typical tangent. Then it is
clear geometrically that
(17.7) PIPO = minimum distance from Po to S,
= maximum distance from Po to a tangent,
= max d.
Furthermore, since d is itself a minimum, namely the mInImum
length of a line joining po to a point on the tangent, we see how a
max-min enters.
18. SOLUTION OF INNER PROBLEM
The Euler equation of the minimization problem on the right-hand
side of (17.6) is
(18.1) u" - rp'(v)j2 = 0, u(O) = c, u'(I) = o.
Writing u = c + W, we have
(18.2)
1 il
U = C + - k(s,t)rp'(v(s ds,
2 0
where k(s,t) is the Green's function of
(18.3) u" = r, u(O) = u'(1) = o.
A simple shows that the minimum value is
(18.4)
1 ilil
K(v) = - k(s,t)rp'(v(srp'(v(t ds dt
4 0 0
+ f[(C - v)rp'(v) + rp(v)] dt.
We can thus write
(18.5) min J(u) = max rp(v).
u 11
The first part yields immediate upper bounds; the second part yields
lower bounds. Judicious choices of trial functions u and v will often
yield quite narrow bounds.
19. FURTHER USES OF QUASILINEARIZATION
The same idea can be used to treat unconventional problems in
control theory and in variational theory. Consider, for example, the
problem of minimizing the nonanalytic functional
(19.1)
J(f) = fll - u(r)1 dr,
over all f subject to
(19.2) (a) 0 ~ f ( t ) ~ M,
(b) f fdt < a,
o ~ t ~ 1,
where u and f are connected by the relation
(19.3) u' = -u + f, u(o) = c.
Writing
(19.4) 11 - ul = max (1 - u)rp,
51
and using a min-max theorem, we can interchange min and max and
obtain an explicit analytic solution of the problem. The details will be
found in the monograph cited in the bibliography at the end of this
chapter.
20. DYNAMIC PROGRAMMING
Let us briefly indicate the use of dynamic programming and, more
generally, invariant imbedding to avoid two-point boundary-value
problems. This is an important consideration since a major source of
error in the application of quasilinearization arises from the solution
of the linear set of algebraic equations associated with the boundary-
value problem. This is no particular problem here, but it can be
significant in higher-dimensional problems.
To illustrate how dynamic programming circumvents two-point
boundary-value problems, consider the problem of minimizing the
quadratic functional
(20.1) J(u) = i
T
[U,2 + b(t
1
)U
2
] dt
1
,
over all u for which the integral exists, satisfying the initial condition
u(a) = c.
The Euler equation is
(20.2) u" - b(t)u = 0, u(a) = c, u'(T) = o.
Let us, however, approach the minimization in a different fashion.
Write
(20.3) min J(u) = f(a,c).
u
Then the principle of optimality yields the nonlinear partial differen-
tial equation
(20.4)
_ of = min [V2 + b(a)c
2
+ v Of],
oa oc
with the initial condition j(T,c) = o. Carrying out the minimization,
we have
(20.5)
- : ~ = b(a)c
2
- HZf
To solve this, we observe that u, by virtue of (20.2), depends linearly
on c, and hence j(a,c) is a quadratic function of c,
(20.6) f(a,c) = r(a)c
2
.
Substituting in (20.5), we see that rea) satisfies the Riccati equation
(20.7) -r'(a) = b(a) - r(a)2, r(T) = o.
This is an initial value problem.
In this case, there is no difficulty using either method. In general,
in multidimensional variational processes, the important point is that
the dynamic programming approach leads to an ordinary system of
differential equations with initial conditions which can easily be
resolved computationally.
21. INVARIANT IMBEDDING
Alternatively, we can regard the missing initial condition in (20.2),
u'(a), as a function of the parameter a. It turns out that this function
also satisfies an ordinary differential equation of Riccati type. This
approach is useful in the more general case when the linear differ-
ential equation under consideration is not the Euler equation of a
variational problem.
22. COMBINATION OF DYNAMIC PROGRAMMING
AND QUASILINEARIZATION
S3
The general pro blem of determining the minimum of the functional
(22.1) J(u) = iT g(u',u,t) dt,
over all functions subject to u(a) = c, may also be approached by
means of dynamic programming. Setting
(22.2) min J(u) = f(a,c),
u
we obtain the nonlinear partial differential equation
(22.3) - of = min [g( v,c,a) + v Of],
oa v OC
with the initial conditionj(T,c) = O. This may be solved computa-
tionally with no difficulty. This approach possesses certain advantages
over quasilinearization, and many other techniques, in that the length
of the interval [a,T] is of slight importance and the presence of
constraints actually simplifies the computational solution.
In higher dimensions, however, the computational solution of
(22.3) is obstructed by rapid-access storage requirements. We can,
to overcome this obstacle, use a combination of local quasilinear-
ization plus global dynamic programming. We shall discuss this in
Chapter Seven.
Chapter Two
1. We have resolutely ignored the very difficult problem area of the existence
and uniqueness of solutions of two-point boundary-value problems. For this,
the reader is referred to
L. COLLATZ, The Numerical Treatment of Differential Equations, Springer,
Berlin, 1960.
4. For an introduction to the application of vector-matrix techniques to
linear systems of differential equations, see
R. BELLMAN, Introduction to Matrix Analysis, McGraw-Hill Book Company,
5. For a discussion of Green's functions, see
E. CODDINGTON and N. LEVINSON, Theory of Ordinary Differential Equations,
McGraw-Hill Book Company, Inc., New York, 1955.
6. A discussion of some of the roles of convexity in analysis is given in
E. F. BECKENBACH and R. BELLMAN, Inequalities, Springer, Berlin, 1961,
together with many additional references.
7-10. We follow the presentation of
and Monotone Convergence," J. Math. Mech., Vol. 8, 1959, pp. 519-574.
16. See
R. BELLMAN and R. KALABA, "Dynamic Programming, Invariant Imbedding
and Quasilinearization: Comparisons and Interconnections," in Computing
Methods i!l Optimization Problems, A. Balakrishnan and L. Neustadt (eds.),
Academic Press Inc., New York, 1964, pp. 135-145.
17. Here we follow
R. BELLMAN, "Quasi-linearization and Upper and Lower Bounds for Variational
Problems," Quart. Appl. Math., Vol. 19, 1962, pp. 349-350.
Further results may be found in
M. A. HANSON, "Bounds for Functionally Convex Optimal Control Problems,"
J. Math. Anal. Appl., Vol. 8, 1964, pp. 84-89.
R. BELLMAN, "Functional Equations and Successive Approximations in Linear
and Nonlinear Programming," Naval Res. Logist. Quart., Vol. 7, 1960,
pp.63-83.
---, "An Approximate Procedure in Control Theory Based on Quasilinear-
ization," Bull. Univ. Jassi, Vol. X(XIV), 1964, fasc. 3-4.
---, "An Application of Dynamic Programming to Location-allocation
Problems," SIAM Rev., Vol. 7, 1965, pp. 126-128.
We are using the concept of duality in a way closely related to that of Fried-
richs; for an application, see
T. IKEBE and T. KATO, "Applications of Variational Method to the Thomas-
Fermi Equation," J. Phys. Soc. Japan, Vol. 12, 1957, pp. 201-204.
19. For a detailed discussion, together with some further applications, see
R. BELLMAN, I. GLlCKSBERG, and O. GROSS, Some Aspects of the Mathematical
Theory of Control Processes, The RAND Corporation, R-313, 1958.
---, "Some Nonclassical Problems in the Calculus of Variations," Proc.
55
20. For various applications of this approach, see
R. BELLMAN and S. DREYFUS, Applied Dynamic Programming, Princeton Univer-
sity Press, Princeton, New Jersey, 1962.
R. BELLMAN, Adaptive Control Processes: A Guided Tour, Princeton University
Press, Princeton, New Jersey, 1961.
21. For invariant imbedding, see
R. BELLMAN, R. KALABA, and G. M. WING, "Invariant Imbedding and Mathe-
matical Physics-I: Particle Processes," J. Math. Phys., Vol. 1, 1960, pp.
280-308.
R. BELLMAN, R. KALABA, and M. PRESTRUD, Invariant Imbedding and Radiative
Transfer in Slabs of Finite Thickness, American Elsevier Publishing Company,
R. BELLMAN, H. KAGIWADA, R. KALAB A, and M. PRESTRUD, Invariant Imbedding
and Time-dependent Transport Processes, American Elsevier Publishing
Company, Inc., New York, 1964.
R. KALABA, "Invariant Imbedding and the Analysis of Processes," Chapter 10
in Views of General Systems Theory, ed. by M. Mesarovic, John Wiley & Sons,
New York, 1964.
Chapter Three
MONOTONE BEHAVIOR AND DIFFERENTIAL
INEQUALITIES
I. MONOTONICITY
Let us now focus our attention upon the monotone properties of
the sequence of approximations obtained using quasilinearization.
The investigation is significant for several reasons. In the first place,
the results are of analytic importance, since they provide new
techniques for establishing the boundedness and convergence of the
approximations and, indeed, as we have noted in connection with the
Riccati equation, occasionally yield the maximum interval of con-
vergence. Secondly, the monotone convergence is quite useful com-
putationally in providing bounds on the solution and in furnishing
a check on the numerical results.
The study of the monotone character of the approximations for
second-order nonlinear differential equations is equivalent to the
study of differential inequalities of the form
(1.1) u" + p(t)u' + q(t)u ~ o.
The study of when this implies that u ~ 0 or u ~ 0 leads us into
a number of classical areas of analysis, some still well tended and
some allowed to fall into neglect. We shall resist the temptation to
make any extensive incursion and restrain our attention to those
results specifically required at the moment. Many references to
further results will be found at the end of the chapter.
In the course of applying quasilinearization to the Riccati equation,
we made repeated use of the fact that we possessed a simple explicit
solution of the first-order linear equation
( 1.2)
u' = a(t)u + bet).
This solution in terms of quadratures and exponentials enabled us to
obtain all the information we desired concerning the differential
inequality
(1.3) u' - a(t)u - b(t) ~ o.
In studying (Ll), we do not possess this luxury and, therefore,
must use some ingenuity. In this chapter we present an assortment
of artifices, devices, and stratagems, ranging from the calculus of
variations to partial differential equations, useful in this area.
Let us note, to begin with, that it is sufficient, in general, to study
the inequality
(1.4) un + q(t)u ~ 0,
since the change of variable, U = rSPdtl2v, which preserves sign,
converts (1.1) into the form
(1.5) v
n
+ (q(t) - p'/2 - p2/4)v ~ o.
1. AN ELEMENTARY APPROACH
Prior to recourse to some high-powered techniques, let us present
a direct elementary method for the study of the inequality
(2.1) un - a(t)u ~ 0, 0 ~ t ~ b,
which can be applied to the study of nonlinear differential equations,
u (t)
--
-....
o I,
Fl 3.1
.......... "
,
,
,
\
\
t, b
MONOTONE BEHAVIOR AND DIFFERENTIAL INEQUALITIES 59
elliptic and parabolic partial differential equations, and other types
of functional equations as well.
Suppose that aCt) > 0 in [O,b] and that u(O) 0, u'(O) > O. We
wish to show that under these assumptions u(t) 0 for 0 < t b.
Suppose that this result were not true so that u(t) had a zero nearest
the origin, say at t = t10 as indicated in Fig. 3.1. There would then
necessarily be an intermediate point t2 at which u(t) possesses a
relative maximum. At this point, however,
(2.2) u"(t) a(t)u(t) > 0,
a contradiction to the statement that u has a local maximum at t = t
2

This establishes the fact that u(O) 0, u'(O) > 0 implies that
u(t) > 0 for 0 < t b.
3. RELATIONS BETWEEN SOLUTIONS AND
COEFFICIENTS
To lay the foundations for our first deeper discussion of the
differential inequality of (2.1), we shall consider some relations
between solutions and coefficients which are analogues of those
relating the roots and coefficients of polynomial equations.
Let U
1
and U
2
be two linearly independent solutions of
(3.1) u" + p(t)u' + q(t)u = 0,
and consider the determinantal equation
u" u' u
(3.2)
Since this is a linear second-order differential equation which
possesses the two linearly independent solutions u = U
1
and u = U2,
it must be equivalent to the original equation.
the notati::."U,) = I U11
U
z
Uz
(the Wronskian of U
1
and U2), we may write (3.1) in the form
l
UI U: I I U: U: I
(3.4) u" + U' U2 U2 + U U
2
U2 = 0
W(t) W(t)
(we write, as before, W(t) == W(U
1
,U2)), a relation which upon
comparison with (3.1) yields the expressions we want for p(t) and
q(t) in terms of U
1
and U2'
Since the presence of W(t) in the denominator is disturbing as far
as possible zero values are concerned, let us now establish the result
of Abel,
(3.5)
From (3.3) we have, upon differentiation,
(3.6)
dW d ( , ') " "
- = - U
1
U
2
- U
2
U
1
= U
1
U2 - U
2
U
1
dt dt
= + q(t)Ul)U2 + + Q(t)U2)
= - = -p(t)W.
This is equivalent to the equation in (3.5). From this it follows that
linear independence of Ul and U
2
at one point, say t = 0, implies
linear independence everywhere. If we take U
1
and U2 to be principal
solutions, u1(0) = 1, = 0, u
2
(0) = 0, = 1, we have
W(u
1
(0),U2(0)) = 1. Hence W(t) is positive for t 0, and, in par-
ticular, not equal to zero, provided that the function f; p(s) ds is
finite for t O. We have already used this result in Sec. 3 of Chapter
Two.
4. A USEFUL FACTORIZATION OF THE SECOND-
ORDER LINEAR DIFFERENTIAL OPERATOR
Let us now perform some simple manipulations which will provide
us with a very useful decomposition of the equation
(4.1) u" + p(t)u' + q(t)u ::0 0,
MONOTONE BEHAVIOR AND DIFFERENTIAL INEQUALITIES
61
or, equivalently, of the operator D2 + p(t)D + q(t), where D = d/dt.
Begin with the relation
(4.2)
J:' ,::1,
dt U
1
U
1
and the similarly derived relation
(4.3)
:t (Iu: 1/1:: I)
I:: 11:
1
:: 1-1:1 :: II:: I

I
Ul /2
U
2
U
2
We suppose that U
1
and U2 are such that W(Ul>U
2
) =F O. Now let us
invoke some simple determinantal results. Let A be a 3 x 3 matrix
(4.4)
with the inverse
(4.5)
where Aii is the cofactor of at; in the determinantal expansion, and
IAI denotes the determinant of A. Since (A-l)-1 = A, we have the
identity
(4.6)
or
I
A22 A321
(4.7) aulAI = A
23
A .
33
Consider now the determinant
(4.8)
u u' u"
Letting Aii denote the cofactor of the ijth element, and using (4.7),
we have
(4.9)
Identifying these cofactors with the terms on the right-hand side of
(4.3), we obtain the relation
(4.10)
Hence,
(4.11)
W(U
I
,U2,U) = W(U
I
,U
2
).!!( W(uI,U)
W(U
I
,U
2
) U
I
dt W(U
I
,U
2
)
W(U
I
,U
2
) d ( u ~ d (u))
= U
I
dt W(U
I
,U
2
) dt U
I
'
or, finally, the desired decomposition
(4.12)
u" + p(t)u' + q(t)u = W(U
I
,U
2
) ~ ( u ~ ~ ( ~ ) ) .
U
I
dt W(U
I
,U
2
) dt U
I
In the following chapter, we will indicate the corresponding result
for the Nth-order linear differential equation.
5. A POSITIVITY RESULT
From the final result of the foregoing section, (4.12), we can easily
demonstrate that a sufficient condition that the relation
(5.1) U" + p(t)u' + q(t)u 20, u(O) = c
I
, u'(O) = c
2
,
in 0 ~ t ~ b, implies that u 2 v, where v satisfies the differential
equation
(5.2) v" + p(t)v' + q(t)v = 0, v(O) = CI, v'(O) = c
2
,
is that there exist a solution of the homogeneous equation
w
N
+ p(t) .... ' + q(t)w = 0
63
which is positive in [O,b]. Call this solution Ul(t), and let ult) be a
linearly independent solution. Then W(U
1
,U2) is nonzero, and we
can choose U2 so that it is positive. As we know, this involves only
a choice of U2(O) and u ~ ( O ) .
From (5.1) and (5.2), we conclude that
(5.3) (u - v)" + p(t)(u - v)' + q(t)(u - v) = ret),
where ret) 2 0 for 0 ~ t ~ b. Applying (4.12), we may write (5.3)
in the form
(5.4)
d ( u ~ d (u - v)) u1r(t)
dt W(U
b
U
2
) dt ~ = W(U
1
,U2) ,
where U
1
and U2 are as indicated above. Hence, integrating between
o and t,
(5.5)
u ~ !!(u - v) =It u1r(s) ds > 0
W(U
1
,U
2
) dt U
1
0 W(U
1
,U
2
) - ,
since, by hypothesis, U
1
and r are nonnegative and W is positive.
There is no constant of integration, since u'(O) = v'(O). Integrating
once again,
(5.6)
Again there is no constant of integration, since u(O) = v(O). From
this we see that u 2 v, the desired result. Furthermore, we have a
lower bound for U - v in terms of L(u) - L(v), where L is the
second-order linear operator. Finally, let us note that the existence
of u1(t) is a characteristic value condition.
6. A RELATED PARABOLIC PARTIAL DIFFERENTIAL
EQUATION
We now wish to study the positivity, or, more precisely, non-
negativity, of the solution of
(6. J) u" + q(x)u + f(x) = 0,
0< x < b,
with the boundary conditions u(O) = u(b) = 0, by means of the
behavior of the solution of the associated heat equation
(6.2)
U
t
= u"'''' + q(x)u + f(x)
with the initial condition
(6.3) u(x,O) = hex),
~ x ~ b,
and the boundary conditions
(6.4) u(O,t) = u(b,t) = 0, t > 0.
We have changed variables in order to keep x as a "space variable"
and t as a "time variable," for the benefit of our intuition.
The motivation for the introduction of (6.2) is the following.
Under appropriate conditions, which we shall examine below, the
limiting value of u(x,t), as t - co, is the solution of (6.1). Hence, if
we can establish the nonnegativity of the solution of (6.2), for
t ~ 0, we will have established the desired non negativity for the
solutions of (6.1). We are using an imbedding technique.
Let us suppose that we have established the requisite existence and
uniqueness theorems and that we are presently only concerned with
the positivity of the solution.
Consider, to begin with, the equation
(6.5)
u
t
= u"'''' + f(x),
and return to the diffusion, or random walk, model which generates
this equation. The corresponding difference scheme is
(6.6) u( x, t + ~ 2 ) = u(x + A, t) ; u(x - A, t) + f(x) ~ 2 ,
where x = 0, A, ... , and t = 0, A2/2, .... We define a function
u(x,t) == u(x,t,A) for all x, t ~ 0 by agreeing to use linear inter-
polation between grid points.
It is immediate from the recurrence relation that u(x,t) ~ 0 for all
x, and t ~ 0, if j(x) ~ 0 and u(x,O) ~ O. Assuming that we can
establish that the limit of the solution of (6.6), the function U(X,t,A),
as 6. --+ 0, is the solution of (6.5), we have demonstrated the non-
negativity of the solution of (6.5).
To handle the more general equation
(6.7) U
t
= U"'''' + q(x,t)u + f(x),
in the case where q(x,t) is nonnegative for x, t 0, we write in
place of (6.6) the relation
(6.8)
( +
6.2) u(x + 6., t) + u(x - 6., t)
u x t - =
'2 2
6.
2
6.
2
+ q(x,t)u(x,t) - + f(x) - .
2 2
Once again, the nonnegativity of u(x,t) is immediate. However,
interestingly enough, the restriction that q(x,t) 0 is inessential,
since we can always perform a change of variable
(6.9)
leading to the equation
(6.10) V
t
= V"'''' + (q(x,t) + k)v + f(x)e
kt

Hence, if we can find a constant k such that q(x,t) + k 2 0 for all
x and t, the proof goes through as before.
Assuming, as above, that there is no difficulty in establishing that
the limit of the solution of the difference equation is the solution of
the differential equation, we have demonstrated:
If
(a) U
t
- U"'''' + q(x,t)u 0,
< x < 1,
t > 0,
(b) u(x,O) 0,
x 1,
(c) u(O,t) = u(1,t) = 0, t > 0,
(d) q(x,t) 2 -k(T) > - 00,
x 1,

for any T> 0,
then u(x,t) 0, x 1, t 0.
A great deal of work has been done in connection with the con-
vergence of the discrete process to the continuous one.
7. CHARACTERISTIC VALUES AGAIN
The interesting question now is to determine why we have an
unrestricted positivity result for the partial differential equation, and
yet a restricted result for the ordinary differential equation. As we
shall see, this corresponds to the actual state of affairs, and is not
merely a reflection on our analytic approach.
Let us suppose that
(7.1) u" + q(x)u + f(x) = 0, u(o) = u(1) = 0,
possesses a unique solution u = uo(x), and let u(x,t) denote the
solution of
(7.2) U
t
- UOl'" - q(x)u - f(x) = 0,
u(O,t) = u(l,t) = 0, u(x,O) = hex), ~ x ~ 1.
If we set
(7.3) u(x,t) = uo(x) + v(x,t),
we see that v satisfies the homogeneous equation
(7.4) v
t
- v",,,, - q(x)v = 0,
v(O,t) = v(l,t) = 0, v(x,O) = hex) - uo(x).
For u(x,t) to converge to uo(x) as t --+ 00, it is necessary that v(x,t) --+ 0
as t --+ 00. Using a Sturm-Liouville expansion of the solution, write
ex)
(7.5)
v(x,t) = ! eAktvk(x).
Each pair, Ak and vix), is determined by the Sturm-Liouville equation
(7.6)
with the vk(x) normalized by the condition fl v ~ ( x ) dx = 1. If all
.0
characteristic values of (7.6) are negative, Ak < 0, then v(x,t) --+ 0 as
t --+ 00. Only in this case does u(x,t) --+ u(x) as t ---->- 00, for all forcing
functions j(x), and therefore only under this constraint on the
characteristic values can we conclude the nonnegativity of lI(x) from
that of u(x,t).
8. VARIATIONAL APPROACH
The study of the differential inequality
(8.1) u" + q(t)u ~ 0,
~ t ~ b,
u(o) = u(b) = 0,
is equivalent to the study of the differential equation
(8.2) u" + q(t)u + f(t) = 0,
~ . t ~ b,
u(o) = u(b) = 0,
for an arbitrary nonnegative function f(t). *
Let us now present a method, entirely different from any of the
three preceding, of studying this equation based upon the fact that
(8.2) is the Euler equation associated with the problem of minimizing
the functional
(8.3) J(u) = f[U
f2
- q(t)u
2
- 2f(t)u] dt
over functions u(t) satisfying the constraints u(O) = u(b) = 0, and
for which the integral exists.
It follows that to demonstrate that the solution of (8.1) is non-
negative whenever f(t) 2 0, it suffices to show that the function
which minimizes J(u) is nonnegative whenever f(t) 2 0.
To begin with, let us require that
(8.4)
7T
2
q( t) < - -
- b
2
'
> 0,
(a characteristic-value condition; see Sec. 9). This condition ensures
that J(u) has a finite lower bound. Standard convexity arguments
show that the lower bound is attained by a unique function. Ref-
erences to these background results will be found at the end of the
chapter.
To show that the minimizing function is nonnegative, let us pro-
ceed by contradiction. Suppose u(t) < 0 in an interval [al>bd lying
We have switched back to t as the independent variable.
u
u (t)
o b
Fig. 3.2
within [O,b], as in Fig. 3.2. Consider then a new function wet) equal
to -u(t) in this interval and to u(t) elsewhere. This introduces a
possible discontinuity into u'(t), but does not affect V-integrability.
This change does not affect the quadratic terms, but diminishes the
contribution of the term involvingj(t). This contradicts the fact that
u(t) is the minimizing function.
Note that this argument shows that there is no need for a solution
with negative values, even if we have not previously established
uniqueness of solution.
9. DISCUSSION
The equation
(9.1) u" + ku + sin 7Tt/b = 0, u(o) = u(b) = 0,
with the solution
-sin 7Tt/b
u = ----'-
k - 7T
2
/b
2
(9.2)
shows that the result is best possible as far as a uniform bound on q(t)
is concerned, which is to say, one independent of f(t). If k > 7T
2
/b
2
,
U has a sign opposite to that of sin 7Tt/b.
The quantity 7T2jb
2
enters as the smallest characteristic value of the
Sturm-Liouville equation
(9.3) U" + AU = 0, u(o) = u(b) = 0,
exhibiting once again the link with characteristic values.
The foregoing argument can be used to show the more general
result that u cannot have more changes of sign than !(t), a variation-
diminishing property.
The solution of (8.1) may be written in the form
(9.4) u = lb k(t,s)f(s) ds,
where k is the Green's function. The preceding arguments establish
nonnegativity of the Green's function, and indeed its variation-
diminishing property. For extensive results concerning variation-
diminishing properties of Green's functions, see the references at the
end of the chapter.
Observe that the argument in Sec. 8 can be applied to a general
variational problem of the form
(9.5) J(u) = f[g(U,U') - 2f(t)u] dt,
where g is even in u and u'.
Finally, let us again point out that we have consistently avoided
the more general form
(9.6)
u" + p(t)u' + q(t)u = 0,
since the change of variable
(9.7)
converts this into the equation
(9.8)
v" + (q(t) _ P';t) _ p 2 ~ t ) v = 0.
10. IMPROVEMENT OF CONVERGENCE ESTIMATES
What is really desired in the theory of differential equations, and
functional equations in general, is a method of successive approxi-
mations which yields rapid convergence to the solution whenever
the solution exists, and throughout its whole domain of existence.
Let us show that the monotonicity results established above enable
us to establish this property for quasilinearization in some simple
cases. A full investigation would take us too far afield into the
general area of differential inequalities.
Consider, for example, the equation discussed in Sec. 12 of Chapter
Two,
(10.1) u" = e
U
, u(O) = u(b) = O.
The sequence of approximations is generated by the relation
(10.2)
with Uo conveniently taken as O. Since e
Un
> 0 for all n; we see from
Sec. 2 that we have monotonicity,
(10.3)
Assuming that we already know that a unique solution exists, we
deduce that the sequence {un}, as a monotone bounded sequence,
necessarily converges. Furthermore, we know from a previous
discussion that it converges to u. Since u is continuous and the con-
vergence is monotone, the convergence must be uniform (Dini's
theorem), a point we have also discussed before.
The positivity of the operator thus provides us with a new technique
for establishing convergence, together with improved estimates for
the domain of convergence.
Generally, given the equation
(10.4) u" = g(u,x), u(O) = u(b) = 0,
with the associated linear equations yielding the approximations
(10.5)
U':"I = g(u",x) + (U,,'I - U")g..(Il,,,X),
71
we see that monotonicity depends upon the establishment of a
bound of the form
(10.6)
If gu is automatically positive, as above, or if it is monotone, convex,
or concave, so that we can draw some conclusions from an examina-
tion of the expression -gu(u1,x) or -giu,x), then more general
results can be stated.
We have wished merely to indicate some of the possibilities of this
approach without delving too deeply into the analytic aspects.
II. DISCUSSION
We have employed three distinct types of approach to study the
behavior of the Green's function of the second-order linear differen-
tial equation
(11.1) u" + p(t)u' + q(t)u + f(t) = 0, u(O) = u(b) = O.
The first, contained in Secs. 2-5, has been direct. The second and
third were indirect approaches, based upon associations of the equa-
tion above with other processes and their associated equations.
Examining these processes in detail, we deduced the required posi-
tivity, and more.
Observe then that we have used imbedding techniques. The original
problem was imbedded within a larger family of problems which
were actually easier to resolve. This idea is fundamental in dynamic
programming and invariant imbedding-techniques which we dis-
cussed briefly in Chapter Two. We shall discuss dynamic program-
ming again in Chapter Seven.
Chapter Three
1. The study of differential inequalities in a systematic fashion dates back to
the work of Caplygin. See, for example,
S. A. CAPLYGIN, New Methods in the Approximate Integration of Differential
Equations, Gostekhizdat, Moscow, 1950. (This is a set of reprints of the
original papers.)
These investigations lead directly into the study of generalized convexity, an
area investigated by Beckenbach, Bing, Valiron, Bonsall, Peixoto, Motzkin,
Tomheim, Curtis, Reid, Hartman, Redheffer, and others. References may be
found on page 160 of
E. F. BECKENBACH and R. BELLMAN, Inequalities, Springer, Berlin, 1961.
See also the books
M. A. KRASNOSELSKII, Positive Solutions of Operator Equations, Gostekhizdat,
Moscow, 1962.
L. COLLATZ, Numerical Treatment of Differential Equations, Springer, Berlin,
1960.
2. Further e x ~ m p l e s of what can be done by arguments of this type may be
found in Chapters 6 and 7 of
R. BELLMAN, Stability Theory of Differential Equations, McGraw-Hill Book
For an application of similar reasoning to harmonic functions, that is, solutions
of u",,,, + u
lIV
= 0, see
E. C. TITCHMARSH, Theory of Functions, Oxford University Press, London, 1939,
p.167.
3. The corresponding results for the nth-order linear differential equations
are given in
G. POLYA, "On the Mean-value Theorem Corresponding to a Given Linear
Homogeneous Differential Equation," Trans. Amer. Math. Soc., Vol. 24,1922,
pp. 312-324.
L. SCHLESINGER, Handbuch der Theorie der linearen Differentialgleichungen,
Vol. 1, 1895, p. 52.
5. This result has been rederived by many authors. It appears to go back to
Caplygin. See, for example,
J. E. WILKINS, "The Converse of a Theorem of Tchaplygin on Differential
Equations," Bull. Amer. Math. Soc., Vol. 53, 1947, pp. 126-129.
6. For a discussion of conditions under which the discrete scheme converges
to the continuous, see
O. E. FORSYTHE and W. WASOW, Finite Difference Methods for Partial Differ-
ential Equations, John Wiley & Sons, Inc., New York. 1960.
73
There are many extensions of these techniques and associated applications.
For example, we can replace (6.6) by
N
u(x, t + .:l.:) = ! ai[u(x + WiA, t) + u(x - Wi A, t)] + ... ,
i=1
and by judicious choice oftheai and Wi obtain a much higher degree of accuracy.
Furthermore, we can apply the same idea to nonlinear partial difference equations
and establish nonnegativity in this fashion. Thus U
t
= uu'" can be approximated
to by
U(X, t + A) = u(x + u(x,t)A, t).
See
R. BELLMAN, I. CHERRY, and G. M. WING, "A Note on the Numerical Integration
of a Class of Nonlinear Hyperbolic Equations," Quart. Appl. Math., Vol. 16,
1958, pp. 181-183.
R. BELLMAN, "Some Questions Concerning Difference Approximations to
Partial Differential Equations," Boll. d'Unione Mate., Vol. 17, 1962, pp.
188-190.
R. BELLMAN, R. KALABA, and B. KOTKIN, "On a New Approach to the Computa-
tional Solution of Partial Differential Equations," Proc. Nat. A cad. Sci. USA,
Vol. 48, 1962, pp. 1325-1327.
T. MOTZKIN and W. WASOW, "On the Approximation of Linear Elliptic Partial
Differential Equations by Difference Equations with Positive Coefficients,"
J. Math. Phys., Vol. 31, 1953, pp. 253-259.
Let us note, in passing, an application of quasilinearization to the study of
solutions of the heat equation,
R. BELLMAN, "Some Properties of Summation Kernels," Duke Math. J., Vol. 15,
1948, pp. 1013-1019.
Here we use the result
max /u(x)/ = lim (i
b
u(x)2n dX)1/2n.
a $ ~ $ b n __
OO
a
8. This approach was first given in
R. BELLMAN, "On the Nonnegativity of Green's Functions," Boll. d' Unione Mate.,
Vol. 12, 1957, pp. 411-413.
For the proof of existence of the minimum, see
R. BELLMAN, I. GLICKSBERG, and O. GROSS, Some Aspects of the Mathematical
Theory of Control Processes, The RAND Corporation, R-313, 1958.
9. See
R. BELLMAN, "On Variation-diminishing Properties of Green's Functions,"
For detailed discussion of variation-diminishing transformations and the concept
of total positivity, see
S. KARLIN, "Total Positivity and Convexity Preserving Transformations,"
Proc. Symp. Pure Math., Vol. VII, 1963, pp. 329-347,
and other references given in the book cited in 1 above.
1O. For generalizations of Dini's theorem to LV-spaces, and even more general
spaces, see
K. VALA, "On Compact Sets of Compact Operators," Ann. Acad. Sci. Fenn. Ser.
A, Vol. 351, 1946, pp. 3-8.
W. MLAK, "Note on Abstract Differential Inequalities and Chaplighin Method,"
Ann. Polon. Math., Vol. X, 1961, pp. 253ff.
---, "Note on Maximum Solutions of Differential Equations," Contributions
to Differential Equations, Vol. 1, 1963, pp. 461-465.
Some papers on differential inequalities of particular interest are
L. COLLATZ, "Aufgaben monotoner Art," Arch. Math., Vol. 3, 1952, pp. 366-376.
---, "Applications of the Theory of Monotonic Operators to Boundary
Value Problems," Boundary Problems in Differential Equations, University of
Wisconsin Press, Madison, Wisconsin, 1960.
R. M. REDHEFFER, "Die Collatzsche Monotonie bie Anfangswertproblemen,"
Arch. Rational Mech. Anal., Vol. 14, 1963, pp. 196-212.
A. I. AVERBUKH, "The Connection Between S. A. Chaplygin's Theorem and the
Theory of Optimal Processes," Avtomat. i Telemeh., Vol. 22, 1961, pp. 1309-
1313.
G. S. JONES, "Fundamental Inequalities for Discrete' and Discontinuous
Functional Equations," J. Soc. Indust. Appl. Math., Vol. 12, 1964, pp. 43-57.
Chapter Fou r
SYSTEMS OF DIFFERENTIAL EQUATIONS,
STORAGE AND DIFFERENTIAL
APPROXIMATION
I. INTRODUCTION
In this chapter, we wish to discuss the application of our tech-
niques to the study of higher-order differential equations and systems
subject to two-point and multipoint boundary-value problems. As
before, we shall first present the quasilinearization procedure which
yields quadratic convergence, and then consider questions of mono-
tonicity.
The computational aspects, particularly the problem of economiz-
ing on rapid-access storage, lead us to develop a novel way of treating
successive approximations. This, in turn, leads us to a new technique
for obtaining the numerical solution of differential-difference equa-
tions and of more general functional differential equations of the
type
(Ll) u'(t) = g(t,u(t),u(h(t))).
We also apply quasilinearization in connection with differential
approximation, a method which permits us to reduce quite compli-
cated functional equations to ordinary differential equations. Ex-
amples of the efficacy of this procedure will be given.
The nonlinear differential equation
(1.2)
u
LV
) = g(u, u', ... , U(N-I)
may be treated directly by means of quasiIinearization. It is usually
more convenient, for both analytic and computational purposes, to
convert it into a system of first-order equations by means of the
introduction of the new variables U
h
U
2
, ,UN' where U
1
=
U, U
2
= u', ... , UN = U(N-I). Then (1.2) becomes
(1.3)
As far as questions of monotonicity are concerned, as we shall
see below, there are advantages sometimes in using one form and
sometimes the other.
We shall omit all proofs of convergence, since they follow the
common pattern which we have already established in Chapter Two.
1. QUASILINEARIZATION APPLIED TO SYSTEMS
Let us then consider the nonlinear system
(2.1) i = 1,2, ... , N,
where the values of the Xi are specified by various conditions at
1= 0 and t = b.
We leave out any I-dependence on the right-hand side merely to
simplify the notation. We can either write
(2.2) i = 1,2, ... , N,
or introduce t as a new dependent variable, denoted by X
N
+
h
and
write
(2.3)
dXN+l == 1.
dt
i = 1,2, ... , N,
SYSTEMS OF DIFFERENTIAL EQUATIONS
77
Since this is a system of the original type of one dimension higher,
we may as well keep the original form.
It is convenient to use vector-matrix notation and write
(2.4)
dx
- = g(x),
dt
where x is the N-dimensional vector with components XI. X
2
, , X
N
,
andg(x)istheN-dimensionalvectorwithcomponentsg
i
(x
1
, x
2
,., x
N
),
i = 1,2, ... ,N. Let the boundary conditions have the form
(2.5)
(x(O),a
i
) = bi'
(x(b),ai) = bi'
i = 1,2, ... , k,
i = k + 1, ... , N,
where the 0i are given vectors, the b
i
are given scalars, and (x,y), as
usual, denotes the vector inner product
N
(2.6)
(x,y) = ! XiYi-
i ~ 1
There is no difficulty in applying the same general technique to the
case where there are nonlinear boundary conditions.
Quasilinearization is readily applied as before. We generate a
sequence of vectors {xn} by means of the linear equations
(2.7)
dx
n
+
1
at = g(xn) + J(xn)(xn+l - x
n
), n = 0,1,2, ... ,
and the boundary conditions
(2.8) (xn+1(0),ai) = bi'
(x n+
1
( b ),ai) = bi'
i = 1,2, ... , k,
i = k + 1, ... , N,
with xO an initial guess. Here J(xn) is the Jacobian matrix defined by
(2.9) J(
n) (Ogi( n n n))
X = - X l' X 2' . , X N ,
oX
i
= (Og;) evaluated at x = xn.
ox,
i,j = 1,2, ... , N,
Prior to the presentation of a numerical example, let us discuss what
is involved in the solution of the linear system.
3. SOLUTION OF LINEAR SYSTEM
Consider the equation
(3.1)
dx
- = A(t)x,
dt
subject to the boundary conditions
(3.2)
(x(O),a
i
) = bi'
(x(b),a
i
) = bi'
i = 1,2, ... , k,
i = k + 1, k + 2, ... , N.
If X(/) represents the solution of the matrix equation
(3.3)
dX
dt = A(t)X,
X(O) = I,
where I is the identity matrix, then the general solution of (3.1) has
the form
(3.4) x = Xc,
where c is a vector. Substituting in (3.2) in order to determine c,
we obtain a system of linear equations
(3.5)
(X(O)c,a
i
) = bi'
i = 1,2, ... , k,
(X(b)c,a
i
) = bi' i = k + 1, ... , N,
or
(3.6)
(c,a
i
) = bi'
i = 1,2, ... , k,
(c,X(b)'a
i
) = bi' i = k + 1, ... , N
(where X' denotes the transpose of X), a set of linear algebraic
equations for the components of c.
The condition for a unique solution of this system is, as is to be
expected, of characteristic value type. One way to ensure that it
holds is to take the interval [O,b] to be sufficiently small.
79
Let us strongly emphasize that the reduction of the computational
solution of the original problem to the numerical solution ofa system
of linear algebraic equations by no means ends the matter. Serious
questions of inaccuracy and instability can easily arise. We shall
discuss some of these below. Furthermore, we shall briefly sketch
some alternative approaches which avoid the solution of linear
algebraic systems. One of these is dynamic programming.
As an example ofthe quasilinearization method applied to a higher-
order equation, let us discuss the numerical solution of
(4.1)
with the boundary conditions
(4.2) u(O) = u'(O) = 0, u"(l) = u"'(l) = o.
U sing the recurrence relation
(4.3)
and the initial approximation uo(x) = 0, and proceeding as above,
we generate Table 4.1. Here k = 6, and an Adams-Moulton integra-
tion procedure was used. The entire calculation consumed about
eight seconds on an IBM-7090.
Table 4.1
x u1(x) u.<x) U3(X) u.(x)
0.0 0.0 0.0 0.0 0.0
0.20312 -0.375572 x 10-
1
-0.394922 X 10-
1
-0.394976 X 10-
1
-0.394976 X 10-
1
0.40625 -0.128544 -0.135598 -0.\35617 -0.\35617
0.60156 -0.242775 -0.256792 -0.256831 -0.256831
0.8469 -0.373308 -0.395724 -0.395786 -0.395786
1.0 -0.501782 -0.532645 -0.532730 -0.532730
5. MULTIPOINT BOUNDARYVALUE PROBLEMS
The technique we have used in the preceding section is equally
applicable to the computational solution of equations of the form
dx
(5.1) - = g(x),
dt
where conditions are imposed at many t-points, for example,
(5.2) i = 1,2, ... , N,
where tl ~ t2 ~ ... ~ tN' Problems of this type arise in a very
natural way in connection with the identification of systems.
In more general fashion, we can treat side conditions of the form
(5.3)
f(X(t), dki(t = bi'
i = 1,2, ... , N,
where the integrals are of Riemann-Stieltjes type, and nonlinear
versions of these relations. Some examples of the computational
treatment of multipoint boundary-value problems will be found in
Chapter Six. Relatively little has been done in connection with the
analytic study of these equations by comparison with their scientific
importance.
6. SUCCESSIVE APPROXIMATIONS AND STORAGE
In the previous pages, we discussed the solution of
(6.1)
dx
- = g(x),
dt
(x(O),ai) = bi'
(x(b),ai) = bi'
i = 1,2, ... , k,
i = k + 1, ... , N,
by means of the successive approximations generated by the relation
(6.2) dX
n
+1 = g(x
n
) + J(xn)(xn+1 - x
n
),
dt
(x
n
+1(O),a
i
) = b.. i = 1,2, ... ,k,
i = k + 1, ... N.
81
We also pointed out how quasilinearization could be used directly
in connection with variational problems.
Let us now inquire into the actual computational aspects of the
foregoing algorithm. In order to use this equation to obtain the
values of X
n
+1(t), given the values of xn(t) at a grid of t-points, t = 0,
... , M, it is necessary to store the entire set of values
k = 0, 1,2, ... , M. This is the usual procedure, when the method
of successive approximations is employed, and is the one used in the
preceding calculations.
This is no serious matter if we are computing the solution of a
single equation of the type shown above in Sec. 4, or even the solution
of a low-dimensional system. But if we wish to treat a system of
dimension 50 or 100 where some accuracy is desired, which means
that each function may require the storage of 1000 values, then lack
of sufficient rapid-access storage can easily block our proposed
solution.
In the next section, we shall present a method which enables us to
overcome this difficulty to some extent. In this connection, we en-
counter very interesting and difficult problems in the storage and
retrieval of information, which have not been studied to any extent.
7. SIMULTANEOUS CALCULATION OF
APPROXIMATIONS
Let us discuss the general problem of determining the numerical
solution of the vector equation
(7.1)
dx
- = g(x)
dt
subject to the linear boundary conditions
(7.2)
(x(O),a
i
) = bi'
(x(b).a
i
) = bi'
i = t, 2, ... , k,
i = k + t, ... , N.
We replace, in the usual way, (7.1) by the sequence of equations
(7.3)
where each Xn is subject to the boundary conditions of (7.2).
Let xo(t) be an initial approximation which requires no uncom-
fortable amount of storage, say a vector whose components are all
constants, polynomials in t, or exponentials. Let xl(t) be determined
from (7.2) and (7.3) with n = O.
Briefly, let us recall how Xl is determined. If Xl is the matrix
solution of
(7.4)
and Pl is the vector solution of
(7.5) Pl(O) = 0,
then the solution of (7.3) has the form
(7.6)
where the vector Cl is determined by the system of N linear equations
(7.7) i = 1,2, ... , k,
i = k + 1, ... , N.
Assuming, as is the case if b is sufficiently small, that this system
has a unique solution, Cl is determined. This yields the full initial
value
(7.8)
Carrying out these operations requires the simultaneous solution
of N2 + N first-order differential equations and the solution of a
set of N simultaneous algebraic equations.
Turning to the equation for X
2
, we must solve the system
(7.9)
dP2 = J(X
I
)P2 + g(x
l
) - J(XI)X
I
,
dt
83
where Xl is the vector determined previously. Instead of storing the
values of XI(t), we adjoin to (7.9) the equations
(7.10) dX
I
= J(XO)XI + g(xo) - J(xo)xo,
dt
where C
I
is the initial value determined by (7.8).
We must now solve (N2 + N) + N simultaneous differential equa-
tions plus N algebraic equations in order to determine the constant
vector C
2
in the representation X
2
= X2C2 + P2, the full initial value
required to determine X
2

Continuing in this way, we see that at the Rth step we solve
simultaneously the differential equations
dX
R
(7.11) - = J(XR_I)X
R
,
dt
XR(O) = I,
dPR = J(XR-I)PR + g(X
R
-
I
) - J(XR_I)XR_b
dt
PR(O) = 0,
Xi(O) = C
i
,
i = 1, 2, ... , R - 1.
This entails the simultaneous solution of N2 + RN simultaneous
differential equations with initial conditions, but no storage of pre-
vious approximations. At each step, we must solve linear algebraic
systems of order N.
8. DISCUSSION
Since the convergence is quadratic, we can expect that R will
seldom exceed 5 or 10 at most. Hence, if N = 2, we must handle at
most 24 simultaneous equations; if N = 4, at most 56 equations;
if N = 10, at most 200 equations. All of these are quite modest
demands in terms of the capacities of current computers.
On the other hand, if N is quite large, say 100, then we modify
our procedure in the following ways. To begin with, in place of
solving the matrix equation
(8.1)
at one stroke, an operation requiring the integration of N2 simul-
taneous equations, together with the N equations for the determina-
tion of x,,, we determine the columns of X
n
+
1
one after the other by
solving the equations
(8.2) y;(O) = e,.,
where e,. is the vector with a 1 in the jth component and zeros else-
where,
o
(8.3)
o
This requires the solution of N + RN simultaneous equations N
times. As usual, we are engaging in a trade of time, which we possess,
for storage, which we do not always possess. We shall discuss another
technique for reducing the growth of the order of the simultaneous
system of differential equations below in connection with differential
approximation.
SYSTEMS OF DIFFERENTIAL EQUATIONS 85
9. BACK AND FORTH INTEGRATION
Let us now briefly mention another approach to the solution of
two-point boundary-value problems which avoids both storage and
the solution of systems of linear algebraic equations.
To illustrate the idea in simplest terms, consider the scalar equation
(9.1) u" = g(u,u',t), u(o) = u(b) = 0,
and suppose that we proceed in the following fashion. Guess a value
for u'(O) and integrate (9.1) numerically with the initial values u(O) =
0, u'(O) = Cl. In this way we obtain values u(b), u'(b}. Ifu(b) = 0, to a
desired degree of accuracy, we stop. If not, we integrate backwards
from t = b with the initial values u(b) = 0, u'(b) = db where d
1
is
the value of u'(b), obtained from the previous calculation. In this
way, we obtain values of u(O), u'(O). If u(O) = 0, to a desired degree
of accuracy, we stop. If not, we integrate forward with the initial
values u(O) = 0, u'(O) = C
2
, where C
2
is the value of u'(O) obtained
from the previous calculation.
It can be shown that this method yields a sequence {cn} which
converges to the correct value of u'(O) in a number of cases. What is
needed is a more sophisticated approach, based perhaps upon a
summability method which enlarges the domain of applicability.
10. DIFFERENTIALDIFFERENCE EQUATIONS
Physical processes involving time delays and retardation effects
lead to differential-difference equations rather than the traditional
differential equations. A typical equation is
(10.1) u'(t) = g(u(t),u(t - 1, t 2 1,
with the function u(t) now specified in an initial interval
(10.2) u(t) = uo(t), O ~ t ~ 1 .
There is no difficulty in applying any of a number of standard
numerical methods available for ordinary differential equations to the
computational solution of equations of this nature. In so doing,
however, we must store the values of the function over an interval of
length 1. Can we avoid this? Once again, there is no difficulty if we
are dealing with systems of low order, but high-order systems can
produce storage difficulties.
II. REDUCTION TO DIFFERENTIAL EQUATIONS
Let us begin by introducing the sequence of functions
(11.1) un(t) = u(t + n), o ~ t ~ 1,
for n = 1, 2, .... Then the single differential-difference equation
of (10.1) may be written as a system of equations
(11.2)
for n = 1, 2, ... , with uo(t) given. The initial conditions are
(11.3)
There are two difficulties to overcome in the simultaneous computa-
tional solution of (11.2). The initial values are unknown, and the
function uo(t) must be stored.
Ignoring the storage of uo(t), a point we shall return to below, let us
indicate how the unknown values of un(O) may be determined. Con-
sider the equation
(11.4)
This may be integrated numerically to yield the value u1(1) = u
2
(O).
We now start all over again and integrate the system
(11.5) u{(t) = g(u
1
(t),U
O
(t,
u ~ ( t ) = g(U2(t),U
1
(t,
which yields the value u
2
(1) = us(O).
u
1
(O) = uo(l),
u
2
(O) = u
1
(1),
87
Continuing in this fashion, at the kth stage we solve the system
consisting of the k equations
(11.6) ui(t) = g(ui(t),Ui_l(t, ui(O) = bi' i = 1,2, ... , k,
where the hi are known values.
The determination of the solution over [O,N] thus requires that a
total of 1 + 2 + ... + N = N(N + 1)/2 equations be integrated al-
together. If N is not too large, this is a fairly routine operation requir-
ing a minimum of storage and not too much time. As we shall see in
the section devoted to differential approximation, we possess methods
for reducing this effort.
12. FUNCTIONAL DIFFERENTIAL EQUATIONS
Let us now apply the same general idea to the functional differential
equation
(12.1) U'(t) = g(t,u(t),u(h(t))).
Equations of this type arise in the construction of realistic mathe-
matical models in a number of fields, ranging from electromagnetic
theory and control theory to respiratory physiology and neuro-
physiology.
We suppose that h(t) :s: t for t 2 0, so that the future is deter-
mined by the past. Furthermore, assume that h'(t) > for t 2
and let h-
1
(t), the inverse function, be denoted by H(t). The function
u(t) is taken to be known in some initial interval [0,t
1
], where tl =
H(O).
Let the sequence {t
n
} be defined recursively by the relation
(12.2) tn = H(tn-l), n = 2, 3, ....
Let H(n)(t) denote the nth iterate of H(t), H(n) = H(HP'-ll), n =
2, 3, .... By virtue of our hypotheses, we observe that the function
H(t) maps the interval [tn-htn] onto [t",t
n
+
l
] in a one-to-one fashion,
and H(k)(t) maps [t n-h( n] onto [t n-1+k,t n+k].
Consider the function
( 12.3)
n = O. 1.2 ...
where s is restricted to the interval [O,t
l
], and H(O)(s) == s. Thus the
values of un(s), for ::;; s ~ tlo are the values of u(t) for tn ::;; t ::;; tn+!.
We have
(12.4)
and the derivative of H(n)(s) can easily be evaluated recursively by
the formula
(12.5)
E... H(n)(s) = H'(H(n-ll(s E... H(n-l\s).
ds ds
Now we set t = H(n)(s), where ::;; s ::;; t
l
Then from the equation
in (12.1) we get
(12.6)
u'(H(n)(s = g{u(H(n)(s,u(H(n-l)(s}
= g(un(s),un_l(s, n = 1,2, ....
Referring to (12.4), we see that (12.6) may be written
(12.7) u ~ ( s ) = [:s H(n)(S)] g(unCs),un_tCs, n = 1,2, ....
Thus (12.1) has been replaced by a system of ordinary differential
equations where s now ranges over a fixed interval [O,t
l
]. We now
proceed to obtain the numerical solution, using the repetitive scheme
of Sec. 11.
13. DIFFERENTIAL APPROXIMATION
If we want the solution over a large t-interval, the foregoing method
can become unwieldy because of the large number of differential
equations that require simultaneous solution. What we would like
to do is start the method all over again at some time t = k, with
the function uk(t) serving as a new initial function. The rub is that we
now have to store the functional values, something we tried to avoid
at the outset.
One way to avoid this problem is to approximate to the function
uk(t) by a function of simple analytic structure, such as a polynomial,
an exponential polynomial or a polygonal function. Thus, for
example, if we approximate to uk(t) by means of a polynomial
M
(13.1) uit) ~ ! ant
n
,
n ~ O
to reproduce the values of uk(t) over an interval [O,T], we need only
store the M + 1 coefficients [ao, ah ... , aM] plus a rule for forming
the polynomial given the coefficients. A similar comment applies
if we use the approximation
M
(13.2) uit) ~ ! ane).n
t
.
n ~ l
Both of these approximations are particular cases of approximating
to a function u(t) by means of a solution v of the linear differential
equation
(13.3) v(N) + a
1
v(N-1l + ... + aNv = 0, v(i)(O) = C
i
,
i = 0, 1, ... , N - 1.
One way of obtaining this approximation is to ask for the values of
a
i
and C
i
which minimize the quadratic functional
(13.4) LT (v - u? dt.
This nonlinear optimization problem, a particular case of a general
control and design problem, can be approached by means of quasi-
linearization. This will be discussed in Chapter Six, as a particular
case of the general design and control problem. Here we will present
a simpler approximation technique which often yields satisfactory
results.
14. SIMPLE VERSION OF DIFFERENTIAL
APPROXIMATION
Suppose that u(t) satisfies an ordinary differential equation of the
form
(14.1) lI(m) = h(II,II' ... ,1I(m-U,t), u(i)(O) = C
i
,
i = 0, 1, ... , m - I,
and we wish to find an approximating linear differential equation of
the form
(14.2)
Let us determine the coefficients ai by the requirement that they mini-
mize the quadratic expression
(14.3)
IT(u(N) + a
1
u(N-1l + ... + aNu? dt,
where u is the function determined by (14.1).
The minimization of the expression in (14.3) leads to the system of
N simultaneous linear equations
(14.4) IT U(i)(U(N) + a1u(N-l) + ... + aNu) dt = 0,
j = 0, 1,2, ... ,N - 1.
For moderate values of N, that is, N ~ 20, the computational solu-
tion provides no difficulty once we have evaluated the integrals
appearing as coefficients. We could, if we so desired, integrate by
parts and reduce the evaluation of these integrals to the evaluation of
the integrals f: (U(i))2 dt. For moderate size N, however, it is more
convenient to proceed directly as follows.
Introduce new variables U
ii
' i,j = 0, 1, ... , N - 1, defined by
the equations
(14.5)
dUii (N-il (j)
-=U u
dt '
Ui;(O) = 0,
and solve these differential equations simultaneously with the original
equation for u, namely,
(14.6)
U(N) = g(u,u', ... ,u(N-ll,t).
The values Uii(T) are the desired coefficients for (14.4).
Having determined the coefficients at by means of the foregoing
procedures we now wish to determine the approximating function
91
vet) as a solution of (14.2). A first approach IS to use the initial
values
(14.7) i = 0, 1,2, ... , N - 1,
and indeed this is what we do below with some success.
In general, however, we would proceed in the following fashion.
Let VI, V
2
, , VN be the N principal solutions of (14.2), the solutions
determined by the condition that the matrix whose columns are
(VI(O), ... , vi
N
-
ll
(O, etc., is the identity matrix.
Every solution can then be written in the form
(14.8)
N
u(t) = L CiVi,

where the C
i
are scalars. Let us choose these Ci so as to minimize the
expression
(14.9)
r T (u _ .f C
i
Vi)2 dt.
Jo
The equations for the C
i
are
i
T N iT
UVi dt - LC
i
ViV
j
dt = 0,
o 0
(14.10) i = 1,2, ... , N.
To determine the various integrals, we introduce as above the vari-
ables Wi and Wij by means of the relations
(14.11)
dWi
- = UVi,
dt
Vi(O) = 0,
dWij
- = vivj '
dt
adjoin the equations for the Vi (Eq. (14.2) with appropriate boundary
conditions) and the equation for u, and integrate.
15. A RENEWAL EQUATION
Let us now discuss the equation of renewal type
(15.1 )
(t
u(t) = J(t) + Jo e-(t-.l u(s) ds.
Taking 0 ~ t ~ 1, we obtain as a third-order differential approxi-
mation to the function e-
t2
a solution of the equation
(15.2) U(3) + 2.740299u(2) + 7.9511452u(1)+ 5.7636455u = o.
Using the initial values obtained from e-
t2
, namely
(15.3) u(O) = 1, u'(O) = 0, u"(O) = -2,
we found such excellent agreement between u(t), the solution of
(15.2), and e-
t2
over 0 ~ t ~ 1, that there was no need to follow the
procedure of Sec. 14.
Consider the expression
(15.4) wet) = fk(t - s)u(s) ds.
Differentiating repeatedly, and adding with the coefficients obtained
above, we have
(15.5) W(3) + 2.740299w(2) + 7.9511452w(l) + 5.7636455w
= k(O)u"(t) + k'(O)u'(t) + k"(O)u(t)
+ 2.740299[k(0)u'(t) + k'(O)u(t)] + 7.9511452k(0)u(t)
+ fU(S)[k"'(t - s) + 2.740299k"(t - s)
+ 7.9511452k'(t - s) + 5.7636455k(t - s)] ds.
Taking k(t) = e-
t2
and assuming that the term under the integral
sign is negligible, we obtain a third-order linear differential equation
for w = u - f
Let us take J = 1 - t e-
s2
ds so that the equation
(15.6) u(t) = 1 - fe-
s2
ds + fe-(t-S)2U(S) ds
has the solution u(t) = 1.
The function J(t) as given above satisfies the linear differential
equation
(15.7)
with
f(O) = 1. 1'(0) = -I. reO) = o.
93
Solving (15.7) together with the approximate linear equation for
u obtained from (15.5), we obtain the values for u(t) shown in Table
4.2.
Table 4.2
1 u(/) u'(/) uH(/)
0.1 0.999999 - -
0.2 0.999999 -0.146 x 10-
3
-0.148 X 10-
2
0.3 0.999969 -
-
0.4 0.999937 -
-
0.5 0.999909 -
-
0.6 0.999898 -
-
0.7 0.999909 0.299 x 10-
3
0.174 X 10-
3
0.8 0.999938 0.330 x 10-
3
0.167 X 10-
3
0.9 0.999970 0.272 x 10-
3
0.135 X 10-
2
1.0 0.999989 0.919 x 10-
4
0.189 X 10-
2
As we can see, the agreement with the desired solution, u(t) = 1,
is excellent.
16. DISCUSSION
Very much the same technique can be used when u(t) satisfies a
differential-difference equation. Alternatively, we can take a set of
t i values and ask that the expression
M
(16.1) L (u(t
i
) - V(t
i
))2
i=1
be minimized over the choice of coefficients and initial values. We
will carry this procedure out in detail in dealing with some identifica-
tion problems in Chapter Six.
Let us also point out that a combination of differential approxima-
tion and quasilinearization can be used to reduce to manageable
terms an integro-differential equation such as
(16.2) u" + g(u) + fk(X - xl)h(u) dX
l
+ r(x) = 0,
u(O) = u(l) - O.
Assume, as is often the case, that k(x) = k( -x) and suppose, for the
sake of simplicity, that the solution of linear equations of differential
approximation yields an approximation of the form (no multiple
roots)
(16.3)
Then
Let
M
k(x) ~ LlXke
Ak
"',
k=l
(16.5) v
k
= e
Ak
'" L"'e-Ak"'lh(U) dx
l
,
Then
(16.6)
and
M
o ~ x ~ l .
viO) = 0,
w
k
(1) = 0,
(16.7) u" + g(u) + L (lk(V
k
+ Wk) + r(x)= 0, u(O) = u(l) = 0.
k=l
We can now apply the standard quasilinearization technique to the
solution of this two-point boundary-value problem.
The reader may, at this point, ask why it is that we did not start with
an approximation of the type appearing in (16.3), rather than the
more complex approach of (13.3). The answer is that there are well-
known difficulties associated with determining the coefficients and
exponents in (16.3). More precisely, simple examples show that the
Ak are unstable functionals of k(x). Essentially, this is due to the
strain on the approximation caused by multiple roots of the charac-
teristic equation of (13.3),
(16.8)
95
On the other hand, the ai' the elementary symmetric functions, are
stable functionals. Consequently, it is better to attempt to determine
the a
i
first, and then the solution of the differential equation.
17. STORAGE AND MEMORY
The preceding discussion indicates quite clearly that we must be
very precise when we speak of storage requirements for the computa-
tional solution of various types of problems. What is always meant,
but, unfortunately, often not explicitly stated, is that a particular
method requires a certain quantity of storage. A more ingenious
method may greatly reduce the requirements.
This is pertinent to the study of the riddle of the human memory,
or memories. When various people remark that the human mind
would require so much energy or so many cells of certain types to
perform various feats, all they mean is that these requirements are
apparently necessary in the light of the few such procedures we
understand or think we understand at the moment. It is quite possible,
and indeed plausible, that in the trial and error process of evolution,
very much more ingenious techniques for storage and retrieval of
information, for pattern recognition, and for decision making have
been developed by living organisms. The discovery of these tech-
niques is an outstanding challenge, and it is certainly rather presump-
tuous to suppose that a few years' effort by a few people will
resolve some of the conundrums of the human mind.
18. MONOTONICITY FOR SYSTEMS
Let us now examine the possibility of obtaining some results
corresponding to those in Chapter Three for systems of differential
inequalities of the form
(18.1 )
dx
- ~ A(t)x,
dt
x(O) = c,
where x is an N-dimensional vector. Here x ;;:: y for two vectors is
shorthand notation for Xi ;;:: Yi' i = 1,2, ... , N. We wish to impose
simple conditions on A(t) which will ensure that x ;;:: y for t ;;:: 0,
where y is determined by the equation
(18.2)
dy = A(t)y,
dt
y(O) = c.
If A(t) is a constant matrix A, we can easily show that a necessary
and sufficient condition for this to be true is that
(18.3) i = j.
The expansion
(18.4)
establishes the necessity, as we see upon examining the right-hand
side in the immediate vicinity of t = O.
Let us establish the sufficiency in the more general case of variable
coefficients by means of the following simple argument. Make the
change of variable
(18.5)
Y
. =
, "
i = 1,2, ... , N,
where the Yi are the components of y and the Zi are the components of
a new vector z. Then (18.2) takes the form
(18.6)
where
(18.7)
with
(18.8)
dz
- = B(t)z,
dt
z(O) = c,
biit) = 0, . i =j,
= a;,(t)eH(a;;-aii) (I", i = j.
If ailt) 0, it follows that all components biJ(t) are nonnegative,
and thus that z(t) z(O) for f O. Hence if c o. y(t) 0 for t O.
Since (18.1) may be written
dx
- = A(t)x + f(t),
dt
x(O) = c, (18.9)
with f(t) ;;::: 0, we see that z = x - y satisfies the equation
(18.10)
dz
- = A(t)z + f(t),
dt
z(O) = O.
97
The same argument as above in (18.5)-(18.8) shows that z ;;::: for
t ;;::: 0.
A result of this type is extremely useful in the study of the system.
(18.11)
dx
- = max [A(q,t)x + b(q)],
dt q
an equation which is basic in the study of Markovian decision pro-
cesses, an important class of dynamic programming processes of
stochastic type.
19. MONOTONICITY FOR Nth-ORDER LINEAR
DIFFERENTIAL EQUATIONS
Consider the linear differential expression
(19.1) L(u) = U(N) + PI(t)U(N-ll + P2(t)U(N-2) + ... + PN(t)u.
Introduce the Wronskian of order n,
fl(t) f;(t)
f2(t) f2(t)
fi
n
-
ll
(t)

fn(t) f<:-l)(t)
Then the analogue of the decomposition for the second-order linear
differential expression, used in Chapter Three, Sec. 4, is valid, namely
(193) Lu - i! ( ... i! ... )
. - Wv I dt W:V2WV dt WI Wa dt WOW2 dt WI .
Here Uh U
2
, , U.v constitute a set of linearly independent solutions
of L(u) = 0.
Let us now introduce what Polya called "property W." A set of
solutions U
1
, U
2
, , UN-l possesses property W in an open interval
(O,b) if
(19.4) U
1
> 0,
in this open interval.
It is clear, using the decomposition of (19.3), that if property W
holds, then
(19.5) L(u) ~ 0, L(v) = 0,
with u(i)(O) = v(i)(O), i = 1,2, ... , kN - 1, implies U ~ v in (O,b).
This is an extensive generalization of the Caplygin result established
in Chapter Three, Sec. 5.
20. DISCUSSION
We have given a few results which enable us to determine when a
sequence of approximations converges monotonically. These results
are not as elegant or as usable as the corresponding results for the
second-order equation, as is to be expected.
In the higher-order case, we are primarily concerned with quadratic
convergence and accuracy. In the next section, we discuss a method
which has proved useful in providing a numerical solution of a linear
differential equation of sufficient accuracy. The importance of this as
far as the effective application of quasilinearization is concerned
cannot be overemphasized.
21. USE OF THE GRAM-SCHMIDT ORTHOGONALIZATION
PROCEDURE IN THE NUMERICAL SOLUTION OF
LINEAR BOUNDARY-VALUE PROBLEMS
There are many pitfalls in the numerical solution of systems of
linear algebraic equations. No method which involyes the solution
of such a system can be considered quite satisfactory. In an earlier
section we showed how we can use a back-and-forth integration
procedure to avoid the numerical solution of linear algebraic systems
during the soluti09 of a linear two-point boundary-value problem.
Alternatively, we may choose the complementary functions (a name
we shall employ for the solutions of the homogeneous differential
equation) in such a manner as to make the solution of the linear
equations for the multipliers as acc;urate as possible. In this section
we shall discuss one possibility in which the Gram-Schmidt orthog-
onalization procedure is employed.
Consider the fourth-order differential equation
(21.1) U(4) - 24u(3) - 169u(2) - 324u
1
- 180u = O.
Its general solution is of the form
(21.2)
Let us focus our attention on the particular solution
(21.3)
which satisfies the initial conditions
(21.4)
u(O) = 0,
u(O) = 0,
U(2\0) = 2,
U(3)(0) = -12.0.
In addition it satisfies the conditions, at t = 1,
(21.5) u(1) = e-
1
- 2e-
2
+ e-
3
= 0.146996,
u(l) = _e-
1
+ 4e-
2
- 3e-
3
= 0.241005 X 10-
1

Let us attempt to determine the function u(t) given in (21.3) as the
solution of (21.1) subject to the boundary conditions u(O) = 0,
u(O) = 0, u(l) = 0.146996, u(l) = 0.0241005.
If we employ the method of complementary functions in a straight-
forward fashion using a Runge-Kutta integration procedure with an
integration interval of 0.01, we obtain disappointing results. Let
wet) be the solution of (21.1) subject to the initial conditions
(21.6) w(O) = 0, W(O) = 0,
and let z(t) be the solution for which
(21.7) z(O) = u, teO) = 0,
Numerical integration yields the values
(21.8) w(1) = 0.195694 X 1010,
w(1) = 0.587083 X 1011,
z(1) = 0.326157 X 10
9
,
t(1) = 0.978472 X 1010.
Using this information to solve the linear algebraic equations
(21.9)
we find
(21.10)
since
w(1)c
1
+ Z(1)C2 = u(1),
w(1)c
1
+ t(1)C2 = u(1),
C
1
= -0.609136 X 10-
2
,
C
2
= 0.365481 x 10-\
C
1
= U(2)(0), C
2
= U(3)(0).
In view of the conditions U(2)(0) = 2, U(3)(0) = -12.0, of (21.4), we
see that we have not succeeded in our task of finding the initial values
of the function u(t).
The difficulty lies in the fact that the characteristic values associated
with the differential operator in (21.2) (-1, - 2, - 3, 30) differ
greatly in their real parts. This means that one of the complementary
functions will dominate the others and make it difficult to determine
what linear combination of complementary functions satisfies the
given boundary conditions. We are encountering the phenomenon
of "ill-conditioning."
101
To remedy the situation we shall pay more attention to which
complementary functions we calculate. Consider the two vectors
oc and (3,
(21.11)
oc = (:,,,) , (3 = (:, .. ) ,
W(3) Z(3)
where the functions w, z, and their derivatives are as above, that is,
they are solutions of the homogeneous equation (21.1) satisfying
the initial conditions in (21.6) and (21.7). Using the Runge-Kutta
integration procedure five times with a grid-size of 0.01, produce the
values of the vectors oc(0.05) and (3(0.05). Next carry out the Gram-
Schmidt orthogonalization procedure on these vectors, leading to
the new orthonormal vectors y(O.05) and b(0.05),
(21.12) (0.05) = oc(0.05)
y I oc(0.05) I '
and
(21.13)
b(0.05) = (3(0.05) - (3(0.05),y(0.05y(0.05) ,
1(3(0.05) - (3(0.05),y(0.05y(0.05) I
where (oc,(3) and loci represent the usual inner product and length.
By carrying out the same linear transformation on the initial vectors
( ~ } ( ~ ) .
we find the new initial vectors which lead to the orthogonal and nor-
malized vectors y(0.05) and b(0.05) via integration of (21.1). These
initial conditions are
(21.14)
and
(21.15)
IP(O.05) _ ( - (P(O.05),y(O.05 ).
Next we integrate (21.1) from t = 0.05 to t = 0.10, using as con-
ditions at t = 0.05, y(0.05) and b(0.05). The vectors so obtained
are orthogonalized and so are the conditions at t = O. Continuing
in this way we eventually find two initial vectors of the form
(D- (fJ
which are initial conditions for (21.1) leading to two orthonormal
vectors at time t = 1. In this way the two-space at t = 1, spanned by
solutions of (21.1) for which u(O) = ti(O) = 0, is accurately deter-
mined. In some instances this will help us to determine the missing
initial values of U(2)(0) and U(3)(0) with precision.
Using this method on the problem (21.1), with the boundary con-
ditions as before, u(O) = 0, u'(O) = 0, u(1) = 0.146996, u'(1) =
0.0241005, we found
(21.16)
For the equation
U(2)(0) = 1.9999997,
U(3)(0) = -12.000003.
(21.17)
U(4) - 9U(3) - 79u(2) - 159ti - 90u = 0,
for which the general solution is
(21.18)
use of the simple method of complementary functions gave the results
U(2)(0) = 2.0013081, U(3) = -12.007849. The boundary conditions
are as above, and the solution is that of (21.3). When the Gram-
Schmidt procedure was used twenty times on the interval from
t = 0 to t = 1, we obtained U(2)(0) = 1.9999995, U(3)(0) =
-12.000003. Again the exact results are given by U(2)(0) = 2,
U(3)(0) = -12.
II. DYNAMIC PROGRAMMING
Let us now indicate how dynamic programming may be used to
avoid two-point boundary-value problems, and the resultant solution
of linear algebraic systems, in treating minimization problems in-
volving quadratic functionals. We have already discussed the scalar
version of this in a previous chapter. Let us here consider the vector
version.
To keep the algebra as simple as possible, let us suppose that we
want to minimize the functional
(22.1) J(x) = LT[(x"X') + (x,Ax)] dt,
subject to the condition x(O) = c. Let A be positive definite so that
no characteristic value problems arise.
The Euler equation is
(22.2) x" - Ax = 0, x(o) = c, x'(T) = 0,
a two-point boundary-value problem. Let us proceed III a quite
different fashion. Write
(22.3) j(c,T) = min J(x).
IX!
Then the usual dynamic programming procedure yields the nonlinear
partial differential equation
(22.4) aj = min [(v,v) + (c,Ac) + (v, gradj)],
aT "
where gradfis the vector whose components are af/aCt, af/aC2, ... ,
aj/ac.\'. and where c .. ('2 , c.v are the components of c.
It is easy to see that the minimum value of v is given by
(22.5) v = -gradff2.
This is the missing initial condition in (22.2).
To make effective use of (22.5), we observe that j(c,T) is a quadratic
function of c. Indeed,
(22.6) f(c,T) = (c,R(T)c).
Using this, we see that
(22.7) v = -R(T)c,
where, using (22.4), R(T) satisfies the Riccati differential equation
(22.8) R(O) = o.
This is an initial value problem which can be solved computationally
in a routine fashion. There are no difficulties as far as R(T) increasing
in magnitude, since we know that R(T) approaches a constant as
T- 00.
A similar result is obtained, with a bit more arithmetic, for the
case where we wish to minimize
(22.9) J(y) = I
T
[(x,AX) + (y,By)] dt
over all y, where
(22.10)
i = Cx + Dy, X(O) = c.
23. INVARIANT IMBEDDING
In many cases we encounter systems of linear differential equations
of the form
(23.1) X' = Ax + By,
y' = Cx + Dy,
x(O) = c,
yea) = d.
If these arise from a variational problem of the type appearing in
105
(22.9) and (22.10), we know how to avoid the two-point boundary-
value problem. If not, we still possess powerful techniques for
finding the missing initial value, y(O), without solving linear algebraic
equations. These techniques are furnished by the theory of invariant
imbedding. They show that if we write
(23.2) y(O) = R(a)c + T(a)d,
then Rand T, as functions of a, also satisfy Riccati-type differential
equations with initial values.
Chapter Four
1. Not much has been done in the area of multipoint problems. For some
recent results and further references see
T. S. VASHAKMADZE, "Multiple-point linear Boundary Problems," AN Gruz
SSR Soobshcheniya, Vol. 35, 1964, pp. 29-36.
2-4. We follow
5. See
R. BELLMAN, H. KAGIWADA, and R. KALABA, "Orbit Determination as a Multi-
point Boundary-value Problem and Quasilinearization," Proc. Nat. Acad. Sci.
USA, Vol. 48, 1962, pp. 1327-1329.
7. These results were given in
R. BELLMAN, "Successive Approximations and Computer Storage Problems in
Ordinary Differential Equations," Comm. ACM, Vol. 4, 1961, pp. 222-223.
R. BELLMAN, R. KALABA, and B. KOTKIN, Some Numerical Results Using
Quasilinearization for Nonlinear Two-point Boundary-value Problems, The
RAND Corporation, RM-3113-PR, April 1962.
9. See
R. BELLMAN, "On the Iterative Solution of Two-point Boundary-value Problems,"
R. BELLMAN and T. A. BROWN, "On the Computational Solution of Two-point
Boundary-value Problems," Boll. d'Unione Mate., Vol. 19, 1964, pp. 121-123.
1O. See
R. BELLMAN, "On the Computational Solution of Differential-difference Equa-
tions," J. Math. Anal. Appl., Vol. 2,1961, pp. 108-110.
R. BELLMAN and B. KOTKIN, "On the Numerical Solution of a Differential-
difference Equation Arising in Analytic Number Theory," Math. Comput.,
Vol. 16, 1962, pp. 473-475.
For a discussion of how differential-difference equations enter into chemotherapy,
see
R. BELLMAN, "From Chemotherapy to Computers to Trajectories," Math-
ematical Problems in the Biological Sciences, American Mathematical Society,
Providence, Rhode Island, 1962, pp. 225-232,
where further references may be found. For a systematic discussion of differential-
difference equations, see
R. BELLMAN and K. L. COOKE, Differential-difference Equations, Academic
Press Inc., New York, 1963.
12. We follow
R. BELLMAN and K. L. COOKE, "On the Computational Solution of a Class of
Functional Differential Equations," J. Math. Anal. Appl., to appear.
R. BELLMAN, J. BUELL, and R. KALABA, Numerical Integration of a Differential-
difference Equation with a Decreasing Time.lag, The RAND Corporation,
RM-4375-NIH, December 1964.
R. BELLMAN, J. BUELL, and R. KALABA, Mathematical Experimentation in Time-
lag Modulation, The RAND Corporation, RM-4432-NIH, February 1965.
13. For further discussion of differential approximation, see
R. BELLMAN, R. KALABA, and B. KOTKIN, "Differential Approximation Applied
to the Solution of Convolution Equations," Math. Comput., Vol. 18, 1964,
pp. 487-491.
R. BELLMAN, B. GLUss, and R. ROTH, "On the Identification of Systems and the
Unscrambling of Data: Some Problems Suggested by Neurophysiology,"
Proc. Nat. Acad. Sci. USA, Vol. 52, 1964, pp. 1239-1240.
---, "Segmental Differential Approximation and the 'Black Box' Problem,"
J. Math. Anal. Appl., to appear.
A. A. KARDASHOV, "An Analysis of the Quality of an Automatic Control
System by the Method of Lowering the Order of the Differential Equation,"
Avtomat. i Telemeh., Vol. 28, 1962, pp. 1073-1083.
Further applications of differential approximation may be found in the book
R. BELLMAN, Perturbation Techniques in Mathematics, PhYsics, and Engineering,
Holt, Rinehart and Winston, Inc., New York, 1964,
107
and in the papers
R. BELLMAN and J. M. RICHARDSON, "On Some Questions Arising in the
Approximate Solution of Nonlinear Differential Equations," Quart. Appl.
Math., Vol. 20, 1963, pp. 333-339.
---, "Renormalization Techniques and Mean-square Averaging-I: Deter-
ministic Equations," Proc. Nat. Acad. Sci. USA, Vol. 47, 1961, pp. 1191-1194.
---, "Perturbation Techniques," Symposium on Nonlinear Oscillations,
Kiev, USSR, 1961.
J. M. RICHARDSON and L. C. LEVITT, ':Systematic Linear Closure Approxima-
tion for Classical Equilibrium Statistical Mechanics-I: General Method,"
Bull. Amer. Phys. Soc., Series II, Vol. 9, 1964, p. 255.
L. C. LEVITT, J. M. RICHARDSON, and E. R. COHEN, "Systematic Linear Closure
Approximation for Classical Equilibrium Statistical Mechanics-II: Applica-
tion to the Calculation of Corrections to the Debye-Huckel Theory of an
Electron Gas," Bull. Amer. Phys. Soc., Series II, Vol. 9, 1964, p. 277.
Some further papers of interest are
L. KH. LIBERMAN, "Problems in the Theory of Approximate Solutions of
Differential Operator Equations in Hilbert Space," lvvz. Matematika, No.3,
1964, pp. 88-92.
E. D. ZAIDENBERG, "A Third Method for the Statistical Linearization of a Class
of Nonlinear Differential Equations," Avtomat. i Telemeh., Vol. 25, 1964,
pp. 195-200.
J. W. CULVER and M. D. MESAROYIC, "Dynamic Statistical Linearization,"
AlEE Joint Automatic Control Conference, New York, June 27-29, 1962.
For a general discussion of the reproduction of functional values, see the
expository article
V. M. TIKHOMIROV, "Kolmogorov's Work on -entropy of Functional Classes
and the Superposition of Functions," Russian Math. Surveys, Vol. 18, 1963,
pp. 51-88,
where many further references may be found.
15. This follows
R. BELLMAN, R. KALABA, and B. KOTKIN, "Differential Approximation Applied
to the Solution of Convolution Equations," Math. Comput., Vol. 18, 1964,
pp. 487-491.
R. BELLMAN and B. KOTKIN, A Numerical Approach to the Convolution Equations
of a Mathematical Model of Chemotherapy, The RAND Corporation,
RM-3716-NIH, July 1963.
16. For examples of the dangers of exponential approximation, see
C. LANCZOS, Applied Analysis, Prentice-Hall Inc., Englewood Cliffs, New
Jersey, 1956.
18. This result is contained in
R. BELLMAN, I. GLICKSBERG, and O. GROSS, "On Some Variational Problems
Occurring in the Theory of Dynamic Programming," Rend. Circ. Mat.
Palermo, Ser. II, 1954, pp. 1-35.
It is interesting to point out that the study of the equation of (18.10), funda-
mental in the theory of Markovian decision processes of continuous type,
focussed attention on this type of functional equation. From this, the results of
R. BELLMAN, "Functional Equations in the Theory of Dynamic Programming-V:
Positivity and Quasi-linearity," Proc. Nat. A cad. Sci. USA, Vol. 41, 1955,
pp. 743-746,
developed, and these, in turn, led to
Finally, let us point out that the study of equations such as (18.10) led to a
natural extension of semigroup theory, embracing the calculus of variations;
see the remarks in
R. BELLMAN, Dynamic Programming, Princeton University Press, Princeton,
New Jersey, 1957, p. 115.
19. See
G. POLYA, "On the Mean-value Theorem Corresponding to a Given Linear
Homogeneous Differential Equation," Trans. Amer. Math. Soc., Vol. 24, 1922,
pp. 312-324.
1. SCHLESINGER, Handbuch der Theorie der Linearen Differentialgleichungen,
Vol. 1, 1895, p. 52, Formula 14.
For extensions to partial differential equations, based upon the work of
Polya and others, see
P. HARTMAN and A. WINTNER, "Mean Value Theorems and Linear Operators,"
Amer. Math. Monthly, Vol. 62, 1955, pp. 217-221.
Further references will be found in
22. For further discussion of these techniques, see
R. BELLMAN, Introduction to Matrix Analysis, McGraw-Hill Book Company,
New York, 1960, Chapter 9,
where many other references may be found. They have been extensively used in
modern control theory by the authors and Adorno, Beckwith, Freimer, Kalman,
Koepcke, and others.
23. For a discussion and application of invariant imbedding, see
R. BELLMAN, R. KALABA, and M. PRESTRUD, Invariant Imbedding and Radiative
109
Transfer In Slabs of Finite Thickness, American Elsevier Publishing Company,
R. BELLMAN, H. KAGIWADA, R. KALABA, and M. PRESTRUD, Invariant Imbedding
and Time-dependent Transport Processes, American Elsevier Publishing
R. BELLMAN, R. KALABA, and G. M. WING, "Invariant Imbedding and Mathe-
matical Physics-I: Particle Processes," J. Math. Phys., Vol. 1, 1960, pp.
280-308.
R. KALABA, "Invariant Imbedding and the Analysis of Processes," Chapter 10
in Views on General Systems Theory, M. Mesarovic (ed.), John Wiley & Sons,
G. M. WING, An Introduction to Transport Theory, John Wiley & Sons, Inc.,
New York, 1962.
For an application of dynamic programming to the numerical solution of
ill-conditioned systems, see
R. BELLMAN, R. KALABA, and J. LOCKETT, "Dynamic Programming and Ill-
conditioned Linear Systems," J. Math. Anal. Appl., Vol. 10, 1965, pp. 206-215.
---, "Dynamic Programming and Ill-conditioned Linear Systems-II,"
These methods are useful in connection with the numerical inversion of the
Laplace transform.
Chapter Five
PARTIAL DIFFERENTIAL EQUATIONS
I. INTRODUCTION
Let us now turn our attention to partial differential equations. As
in the previous chapters, we shall first consider some of the com-
putational aspects and then briefly discuss questions of monotonicity
of approximation and positivity. This chapter is on a different level
from any of the preceding or subsequent chapters, since it presumes a
knowledge of the highly developed and complex art of solving partial
differential equations with a digital computer. The reader unversed
in this cabalistic computing is advised to skip this chapter.
Our first numerical example will be a nonlinear parabolic partial
differential equation and our second a nonlinear elliptic partial
differential equation.
The computational situation here is far more complex than in the
case of ordinary differential equations. In obtaining the numerical
solution of ordinary differential equations, subject to initial value
conditions, we can use methods of such guaranteed accuracy that
we can be quite confident that the results reflect the efficiency of the
technique employed. It is therefore meaningful to compare different
methods. In solving partial differential equations, the problem is
that of trying to determine the contribution of the actual numerical
procedure. As a consequence of the effects of truncation errors,
round-off errors, ofthe word-length and actual computing procedures
of the computer that is used, of the use of various crude bounds on
higher-order derivatives in the choice of techniques, and of the labor
involved in carrying out higher-order difference techniques, it is
not easy to obtain valid a priori estimates of the advantages of higher-
order methods as opposed to lower-order methods, nor to compare
III
different procedures used by different people. It follows that it is
essential to do a certain amount of experimentation.
2. PARABOLIC EQUATIONS
We shall begin with a discussion of parabolic partial differential
equations, since we have already discussed them to some extent in
connection with the positivity of linear differential operators of the
second order.
Given the problem of obtaining a numerical solution of a partial
differential equation of the form
(2.1)
with initial values of the form
(2.2) U(X,O) = hex), 0< x < 1,
and boundary conditions of the form
(2.3) U(O,t) = u(l,t) = 0,
following our usual quasilinearization approach, we set up the
approximation scheme
(2.4)
(U
n
+1)t = (u
n
+1)"", + g(um(u
n
),,)
+ (u
n
+1 - un)g,,(um(u
n
),,)
+ u
n
+1)" - (un)")gux(um(u
n
),,,)
Starting with an assigned value of uo(x,t), we have a linear equation
for each Un> n ~ 1.
Let us then proceed directly tQ the actual numerical aspects.
3. A SPECIFIC EQUATION
Let us consider the equation
(3.1) u
t
- u"'" = (1 + u
2
)(1 - 2u),
which was constructed so as to have the simple explicit solution
(3.2) u(x,t) = tan (x + t).
PARTIAL DIFFERENTIAL EQUATIONS 113
This solution has the additional advantage of becoming infinite as
x + t --+ 7T/2 = 1.57 +, and thus putting numerical algorithms to a
good test. The usual Picard iteration method
(3.3)
was compared with quasilinearization over two triangles 0 :::;; t :::;;
(1 - x), 0 :::;; x :::;; 1, and 0 :::;; t :::;; 1.5 - x, 0 :::;; x :::;; 1.5.
The numerical values were obtained by using conventional differ-
ence approximation techniques which we will describe below. As
mentioned before, other computational algorithms for the solution
of partial differential equations could yield different estimates of the
value of one approach as compared with the other.
4. DIFFERENCE APPROXIMATION
The first question is that of deciding upon a difference approxima-
tion to the terms u
t
- U
XX
' Since the increments = = 0.01
were used, the Crank-Nicholson difference operator was employed
to obtain a discrete system.
Letting um,n denote the solution of the difference equation analogue
of (3.1) at the lattice point nM, we replace (3.1) by
(4.1)

1 {[ (k+1) 2 (kH) (k+1)] [(k
n
) (k
n
) (k
n
)
- u m- I ,n+1 - u m,n+1 + um+I,nH + urn-I,n - 2um,n + um+I,n]f
= +
where feu) = (1 + u
2
)(1 - 2u), and k
n
denotes the final number of
iterations required to obtain an acceptable approximation to the
value of um,n at the grid-points on the line T =
The criterion for acceptance of an approximation was a commonly
used one, namely that the maximum relative change per line be less
than a prescribed amount before passing to the next line. In this case,
we required that
(4.2)
I
U(k) U (k-I) I
m tn,lI - tn,lI .,.., 10-
8
ax (k) ..::::. '
m tim."
5_ COMPARISON BETWEEN PICARD AND
QUASILINEARIZATION
Table 5.1 gives the number of iterations required to achieve the
criterion of (4.2) before continuing to the next t-line.
The significant observation was that while the Picard procedure
and the quasilinearization procedure required about the same amount
Table 5.1
Number of
Base of Triangle Method Iterations I-interval
1.0 Newton 2 (0.01, 0.98)
1 (0.98, 0.99)
1.0 Picard 4 (0.01, 0.68)
3 (0.68, 0.93)
2 (0.93, 0.99)
1.5 Newton 3 (0.01,0.23)
1.5 Picard 28 (0.01,0.50)
of work in the case of the triangle 0 ~ t ~ 1 - x, 0 ~ x ~ 1, about
nine times the number of iterations required for the quasilinearization
method were needed to obtain comparable accuracy by the Picard
procedure when the region of interest was the larger triangle 0 ~ t s
1.5 - x, 0 ~ x ~ 1.5.
6. GRADIENT TECHNIQUES
In Chapter One we indicated that one approach to the determina-
tion ofa solution of the scalar equationf(x) = 0 is to use the Newton-
Raphson approximation method. In place of the original static
process, we consider a discrete dynamic process
(6.1 )
a nonlinear difference equation. The limit of the sequence {x
n
},
under suitable conditions, is the required root.
In place of a discrete process, we can consider a continuous process
based upon the differential equation
(6.2)
dx
dt = f(x),
or, perhaps, in closer analogy with (6.1), the differential equation
(6.3)
dx - f(x)
-=--
dt f'(x)
It is clear that there are arbitrarily many differential equations whose
limiting form as t -+ 00 is the static equation f(x) = O. Which one
we choose depends upon consideration of accuracy, stability, time,
and type of computation available.
The same idea can be used to study general functional equations.
Consider the two-point boundary-value problem
(6.4) u" = g(u,u',x), u(O) = u(b) = O.
Associate with this equation a time-dependent process
(6.5)
u
t
= u"'X - g(u,u""x),
u(O,t) = u(b,t) = 0,
u(x,O) = w(x), 0< x < b,
where w(x) is chosen in some expeditious fashion.
Applying the quasilinearization approach of the preceding sections,
we have a way of studying the solution of (6.4) in a different fashion,
one that mayor may not involve the solution of linear systems
of ordinary differential equations, or linear systems of algebraic
equations.
Let us point out that it may be preferable to use a different
approach to (6.5) than that described above. For example, poly-
nomial approximation techniques appear to be quite effective.
Secondly, let us note that we are not particularly interested in the
complete time history of u(x,t), but only its "steady-state,"
(6.6) u(x) = lim u(x,t),
t-+ 00
hopefully the solution of (6.4). Since we expect this approach to be
exponential, that is,
(6.7) u(x,t) '" u(x) + u
1
(x)e",I<,,)t + u
2
(x)e",2(,,)t + ... ,
we can profitably use nonlinear summability to obtain excellent
extrapolation of u(x) from u(x,t) for small t.
Finally, let us note that (6.5) affords only one method for imbed-
ding (6.4) within a continuous family of processes. The theory of
invariant imbedding, using transport theory models, furnishes
alternative approaches.
These are techniques which would take us too far afield at the
moment. In subsequent volumes we shall return to them. The impor-
tant point to stress is that a mixture of techniques is highly desirable,
and it is unwise and unscientific to attempt to attack significant prob-
lems with only a single method. References to the preceding
approaches will be found at the end of the chapter.
7. AN ELLIPTIC EQUATION
Corresponding experiments were carried out for the elliptic
equation
(7.1)
an equation of interest in magnetohydrodynamics. The boundary
conditions will be described below.
We shall compare the solution obtained by means of the recurrence
relation
(7.2) (u
k
+1)"" + (u
k
+1)YY = eUk,
the Picard approximation technique, with that obtained from
(7.3) (U
k
+
1
)",,,, + (u
k
+1)YY - e
Uk
u
k
+1 = e
Uk
(1 - Uk),
the quasilinearization technique.
There is no difficulty in supplying the analogues of the monotonicity
and quadratic convergence proofs, following the lines indicated in
earlier sections. The required positivity is easily obtained by following
the method of Chapter Three, Sec. 8,. the variational approach.
8. COMPUTATIONAL ASPECTS
We shall consider the rectangular region 0 x !, 0 y 1,
for two sets of boundary conditions, U = 0 on the border of the
rectangle, and U = 10. As is to be expected, the quasilinearization
approach appears in a most favorable light for U = 10, where larger
gradients occur.
The conventional approach was followed. The Laplacian is re-
placed by the difference expression
(8.1) (um+l,n + um,n+l + um-I,n + um,n-I - 4u
m
,n)/h
2
,
where xm,n is the discrete analogue of and h = =
which in this case was taken to be i"4'
If we order the points (m,n) so that (m,n) precedes (m',n') if
n < n' or if n = n' and m < m', and denote the resulting set of
15 x 31 = 465 numbers {Umn,k} by Uk' both the Picard and quasi-
linearization techniques lead to the problem of solving large systems
of linear algebraic 'equations.
Normalizing these equations by dividing through by the negative
of the coefficient of the central term umn,k' we obtain in each case a
system which has the form
(8.2)
Here I, L
k
, and Uk are 465 x 465 matrices; I is the identity matrix,
while Lk and Uk are respectively lower and upper triangular matrices
appropriate to the approximation method used. In our problem,
they are transposes of each other and have only two diagonal lines
of nonzero elements; in the Picard iteration these elements are
identically equal to t, while in the Newton procedure they are equal to
(4 + h
2
exp uk)-t,
where the approximation Uk is evaluated at the central point (m,n) of
the star of points indicated in (8.1). The components of the vector b
k
in (8.2) are equal to the appropriate values of the right-hand sides of
(7.2) and (7.3), respectively.
Since in many realistic studies the size of the problem is so huge
as to preclude the use of Gaussian elimination to solve the algebraic
systems (8.2), we chose the successive over-relaxation method of
Young, primarily because it is the simplest iterative method which is
substantially better than the Gauss-Seidel method although it is
recognized that faster converging block relaxation and alternating
direction methods are not much more complicated than successive
over-relaxation.
Applying the successive over-relaxation method directly to the
systems (8.2) for a fixed value of k would yield the iterative formula
(8.3)
(I - wL
k
)ur+l,k+1
= [WUk - (w - I)I]u
r
k
+1 + b
k
, r = 0, 1,2, ... ,
where w is the over-relaxation parameter and the component
equations are solved successively in the order that produces the
components of the vector U
r
+
1
,k+l consecutively in the order indicated
above. However, it is clear that not much effort should be expended
in obtaining a very accurate solution to (8.2) if U/r+l is not a good
approximation to the solution of (7.1). Similarly, there is not much
point to iterating on the index k if the approximations given by (8.3)
are too rough as a consequence of terminating the iteration on r too
soon. Thus there exists the open problem as to when one should
iterate on k or on r at each step. We avoided the attempt at this
relatively difficult analysis and simply iterated on each simultaneously,
using the formula
(8.4)
the component equations being solved in the order indicated above.
Thus we see a strong interweave between the numerical method used
for solving either (8.2) or (8.4) and the rate of convergence of the
actual solutions of (8.2) or of (8.4), the difficulty of whose analysis
dictates experimentation. We also note that this blend of iterative
procedures for the solution of the linear systems (8.2) and the Newton
or Picard iterations to solve the original nonlinear problem has the
advantage that one does not have to go to the full-scale effort of
solving accurately the system (8.2) for each value of k.
As in the parabolic case the criterion for stopping the iterations
was that the maximum of the absolute value of the relative change
between consecutive iterations of the functional values be no greater
than 10-
6
If this were truly a bound on the relative error of the
numerical solution and the solution of nonlinear equation (7.1) with
the Laplacian replaced by the expression (8.1), then this criterion
would be about two orders of magnitude too stringent to be justified
by the truncation errors. However, some measure of stringency is
dictated by the heuristic character of the stopping criterion when one
knows neither the true solution nor effective bounds on it.
In the case of Picard iteration where Lk and Uk are independent of
k, if one were to ignore the fact that b
k
depends on k, the theoretically
optimal value of ill is 1.732 for the fastest convergence. For the
Newton iteration the value would be a little smaller, again ignoring
the change in Lk and in Uk as well as in b
k
from iteration to iteration,
which we clearly cannot do. However, in the case of zero boundary
conditions for the Picard iteration, where Lk and Uk are constant, the
best value (or values) of ill are undoubtedly near the theoretical value
noted above because the change in Uk over the region is very small
(the value at the center of the region obtained for the final solution
of 0.007071 to four significant figures). Consequently, for this case
we experimented only with the values ill = 1.70 and 1.74. No such
confidence existed for the case of Picard iteration on the situation
where U is equal to 10 on the boundary because of the large variations
in Uk over the region, which would undoubtedly introduce large
changes at a point from iteration to iteration. Of course, in the
Newton iterations, since Lk and Uk change from iteration to iteration,
one has even less of an idea as to the optimal value of ill. Consequently t
in the remaining cases a number of runs for different values of w
were made to estimate an optimal value of w. Table 5.2 summarizes
the results.
From the table we observe that for the case of zero boundary
conditions no advantage was obtained from the use of the quad-
ratically converging Newton procedure over the results obtained
Table 5.2
Value of U(k)
Value of w Number of Iterations at (t, i)
u == 0 (Picard Method)
1.74 64 -0.007071
1.70 77 -0.007071
u == 0 (Newton Method)
1.95 258 -0.007071
1.74 64 -0.007071
1.70 77 -0.007071
1.50 154 -0.007071
1.00 426 -0.007071
u == 10 (Picard Method)
1.55 63 5.65995
1.50 53 5.65995
lAO 66 5.65994
1.35 71 5.65993
1.30 76 5.65993
1.20 78 5.65992
1.10 106 5.65997
1.00 134 5.65998
u == 10 (Newton Method)
1.74 58 5.65995
1.70 51 5.65995
1.65 46 5.65995
1.60 42 5.65995
1.55 41 5.65995
1.50 50 5.65995
lAO 66 5.65996
1.00 148 5.65998
from the first-order converging Picard procedure. In fact, the same
values of w gave the most rapid convergence in both cases. On the
other hand, when the boundary conditions were raised to 10, implying
rapid changes in U near the boundary, the Newton procedure was
more advantageous, requiring some 41 iterations as compared with
53 for Picard's. We note further that curiously the optimal value of
w was slightly lower for Picard's procedure.
As an additional note we included the case of Gauss-Seidel
relaxation given by w = 1 for comparison with successive over-
relaxation. As expected, the successive over-relaxation method ranged
from two and one-half to almost seven times faster than Gauss-
Seidel relaxation.
Another incidental observation was that the points at which the
maximum relative change in Uk took place were invariably very close
to the origin. In most cases it was at the point (i4",i4")'
9. POSITIVITY AND MONOTONICITY
Any detailed discussion of positivity and monotonicity for partial
differential operators would enmesh us too deeply in the theory of
partial differential equations. Let us merely point out a few simple
consequences of the techniques we have already applied.
The method of finite differences, used in Chapter Three, Sec. 6,
readily establishes the desired nonnegativity of the solution of
(9.1)
U
t
= U"'''' + f(x,t), t> 0,
U(x,O) = g(x),
~ x ~ 1,
U(O,t) = u(l,t) = 0,
t> 0,
where f, g ~ 0, and for similar equations in higher dimensions.
To obtain the corresponding result for elliptic partial differential
equations such as
(9.2)
U
zz
+ U
NN
+ q(x,y)u = f(x,y).
we can either use the parabolic partial differential equation as in
Chapter Three, Sec. 6, or we can start from the variational problem
of minimizing the functional
(9.3) flU! + u! + q(x,y)u
2
- 2f(x,y)u] dA
and proceed as before. In this fashion, we can obtain not only the
nonnegativity of the Green's function, but also analogues of the
variation-diminishing property.
10. THE HOPF-LAX EQUATION
Consider the conservation relation
(10.1) U
t
+ f(u)", = 0,
with the initial conditions u(x,O) = g(x), T ~ t ~ O. The case f(u) =
u
2
j2 is particularly interesting since the equation is a limiting case
of the Bateman-Burgers equation
(10.2)
an equation which plays an important role in various simplified
mathematical models of turbulence.
Let us assume that f(u) is strictly convex. Write (10.1) in the form
(10.3)
and introduce the function U(x,t) by means of the relation U", = u.
Then equation (l0.3) becomes
(10.4)
or
(10.5)
upon integrating with respect to x and choosing appropriate initial
conditions.
PARTIAL DIFFERENTIAL EQUATIONS
The associated linear equation is
(10.6) W
t
+ /(V) + (W", - V)/'(V) = 0, W(x,O) = G(x).
Using the easily demonstrated positivity of the operator
a/at + I'(v) a/ax,
we have an explicit solution of (10.5) in the form
(10.7) U = min w(x,t,v),
v
where w is the solution of (10.6).
123
Taking v to be constant, we obtain a simple explicit solution which
is the representation found previously by Lax in the general case, and
Hopf in the particular case where j(u) = u
2
/2, the equation of (10.2).
In this case, the equation can be reduced to the linear heat equation
and solved explicitly.
Chapter Five
1. The numerical results were obtained in conjunction with M. Juncosa. See
R. BELLMAN, M. JUNCOSA, and R. KALABA, "Some Numerical Experiments
Using Newton's Method for Nonlinear Parabolic and Elliptic Boundary-
value Problems," Comm. ACM, Vol. 4, 1961, pp. 187-191,
where further references are given.
6. For a detailed discussion of gradient techniques, see
P. ROSENBLOOM, "The Method of Steepest Descent," Proc. Symp. Appl. Math.,
Vol. VI, 1956, pp. 127-176.
For a brief account of the application of nonlinear summability techniques, see
R. BELLMAN and R. KALABA, "A Note on Nonlinear Summability Techniques
in Invariant Imbedding," J. Math. Anal. Appl., Vol. 6, 1963, pp. 465-472.
For a discussion ofthe use of polynomial approximation in the numerical solution
of partial differential equations, see
R. BELLMAN, R. KALABA, and B. KOTKIN, "On a New Approach to the Com-
putational Solution of Partial Differential Equations," Proc. Nat. Acad. Sci.
USA, Vol. 48, 1962, pp. 1325-1327.
R. BELLMAN, "Some Questions Concerning Difference Approximations to
Partial Differential Equations," Boll. d'Unione Mate., Vol. 17, 1962, pp. 188-
190.
9. For a number of references to positivity properties of the solutions of
partial differential equations, see
These properties are, in turn, closely associated with "maximum principles"
and any detailed discussion would plump us in the middle of the modern theory
of partial differential equations.
1O. See
P. LAX, "Hyperbolic Systems of Conservation Laws. II," Comm. Pure Appl.
Math., Vol. 10, 1957, pp. 537-566,
R. BELLMAN, J. M. RICHARDSON, and S. AZEN, "On New and Direct Compu-
tational Approaches to Some Mathematical Models of Turbulence," Quart.
Appl. Math., to appear.
In order to avoid extensive excursions, we omitted any discussion of the
connection between the minimization of functionals of the form
J(u) = SoT h(u,u') dt
and partial differential equations of the form Ut = g(u,u""t). See
E. D. CONWAY and E. HOPF, "Hamilton's Theory and Generalized Solutions
of the Hamilton-Jacobi Equation," J. Math. Mech., Vol. 13, 1964, pp. 939-
986.
This in turn would lead to the use of approximation in policy space in the
calculus of variations.
For an application of dynamic programming to the study of parabolic partial
differential equations, see
W. H. FLEMING, "The Cauchy Problem for Degenerate Parabolic Equations,"
J. Math. Mech., Vol. 13, 1964, pp. 987-1008. See also Chapter 4 of
Chapter Six
APPLICATIONS IN PHYSICS, ENGINEERING,
AND BIOLOGY
I. INTRODUCTION
In this chapter we wish to sketch some typical applications of
quasilinearization to characteristic problems in physics, engineering,
and biology. We shall discuss in turn problems in optimal design
and control, inverse problems in radiative transfer, questions of orbit
determination, system identification and adaptive control, periodo-
gram analysis, and, finally, an inverse problem in cardiology.
1. OPTIMAL DESIGN AND CONTROL
A basic problem in modern control theory is that of minimizing
the scalar functional
(2.1) ley) = iT g(x,y) dt,
over M-dimensional vector functions y(t), where x, an N-dimensional
vector, and yare connected by the differential equation
(2.2)
dx
- = h(x,y),
dt
x(O) = c.
Frequently, the problem is a more general one of minimizing the
scalar functional
(2.3) l(y,a) = iT g(x,y,a) dt + cp(a)
over M-dimensional vector functions y(t) and K-dimensional vectors
a, where x, y, and a are connected by the differential equation
(2.4)
dx
- = h(x,y,a),
dt
X(O) = c.
' ' ' ~
These entail questions of design since a may be viewed as a vector
of design parameters. We shall not discuss any of the still more
general problems involving functional differential and partial
differential equations.
To treat this variational process by means of quasilinearization,
we can proceed, as we know, in two different fashions. We can begin
with an initial approximation Yo, ao, and generate Xo by means of (2.4),
(2.5)
dxo h(
- = xo,Yo,ao),
dt
xo(O) = c.
To obtain Xl> YI, aI, we consider the equation
(2.6)
where hI is the expansion of h(x,y,a) around xo, Yo, a
o
, up to linear
terms, and pose the problem of minimizing
(2.7) LT g2(xI'YI,aI,xo,Yo,ao) dt
over YI and at> where g2 is the expansion of g(x,y,a) around Xo, Yo, a
o
up to quadratic terms.
It is easy to see that the variational equations are linear in XI. Yl>
and a
l

Alternatively, we can use standard variational techniques to obtain
the Euler equation and then apply quasilinearization to this equation.
To unify the treatment, we can dispense with the fact that a is a
constant by taking it to be a solution of the differential equation
(2.8)
da _ 0
- ,
dt
with an unknown initial condition.
Introducing the Lagrange multipliers A(t) and /-t(t), we then
consider the new criterion functional
(2.9) LT[g(x,y,a) +(,\(t),x - h(x,y,a + (/L(t),d)] dt + tp(a(O,
and write down in standard manner the appropriate differential
equations and boundary conditions.
APPLICATIONS IN PHYSICS. ENGINEERING. AND BIOLOGY 127
3. AN EXAMPLE
Consider, for example, the problem of minimizing the functional
(3.1) J(y,a) = - (x
2
+ /) dt + - ,
1 il a
2
202
over the function y and the parameter a, where
(3.2)
dx
- = -ax + y,
dt
x(o) = c,
da = 0.
dt
The Euler equations and boundary conditions are
(3.3)
x = -ax + y, x(o) = c,
~ = x + Aa, A(I) = 0,
d = 0, {teO) = a(O),
it = AX, {t(1) = 0,
y = A.
The relation y = A enables us to eliminate A from consideration.
As our initial approximation, we set
(3.4) xo(t) = 1,
YoCt) = 0,
{to(t) = 0, ao = 0.
Then, either of the two procedures discussed in Sec. 2 yields the
following scheme of successive approximations:
Yn+l(l) = 0,
There is no difficulty in applying standard techniques discussed
previously to the numerical determination of the sequence {xmy,.,a
n
}.
In the following section, we describe some specific calculations.
T
O.
0.1000
0.2000
0.3000
0.4000
0.5000
0.6000
0.7000
0.8000
0.9000
1.0000
T
O.
0.1000
0.2000
0.3000
0.4000
0.5000
0.6000
0.7000
0.8000
0.9000
1.0000
T
O.
0.1000
0.2000
0.3000
0.4000
0.5000
0.6000
0.7000
0.8000
0.9000
1.0000
Table 6.1
SOME COMPUTATIONAL RESULTS
Iteration 1
X Y MU
0.100000 X 10
1
-0.662196 x 10 0.283436 x 10
0.910758 x 10 -0.566718 x 10 0.223168 x 10
0.830158 x 10 -0.479745 x 10 0.170912 x 10
0.757867 x 10 -0.400410 x 10 0.126965 x 10
0.693161 x 10 -0.327919 x 10 0.906022 X 10-
1
0.635393 x 10 -0.261546 x 10 0.611771 X 10-
1
0.583983 x 10 -0.200628 x 10 0.381112 X 10-
1
0.538418 x 10 -0.144555 x 10 0.208900 X 10-
1
0.498242 x 10 -0.927650 X 10-
1
0.905750 X 10-'
0.463053 x 10 -0.447402 X 10-
1
0.221157 X 10-'
0.432498 x 10 0.167638 X 10-
7
0.125729 X 10-
7
Iteration 2
X Y MU
0.100000 X 10
1
-0.640197 x 10 0.252000 x 10
0.917906 x 10 -0.557950 x 10 0.174601 x 10
0.845465 x 10 -0.481894 x 10 0.128792 x 10
0.781899 x 10 -0.410979 x 10 0.925063 X 10-
1
0.726531 x 10 -0.344449 x 10 0.640778 X 10-
1
0.678772 x 10 -0.281596 x 10 0.421680 X 10-
1
0.638114 x 10 -0.221749 x 10 0.257069 X 10-
1
0.604125 x 10 -0.164271 x 10 0.138468 X 10-
1
0.576444 x 10 -0.108548 x 10 0.592470 X 10-'
0.554777 x 10
-0.539856 X 10-
1
0.143373 X 10-'
0.538896 x 10 0.223517 X 10-
7
0.814907 X 10-
8
Iteration 3
X Y MU
0.100000 X 10
1
-0.640414 x 10 0.232393 x 10
0.917788 x 100 -0.558586 X 10 0.175557 x 10
0.845306 x 10 -0.482596 x 10 0.129699 x 10
0.781742 x 10 -0.411697 x 10 0.933488 X 10-
1
0.726424 x 10 -0.345141 x 100 0.648322 X 10-
1
0.678770 x 10 -0.282227 x 10 0.428066 X 10-
1
0.638276 x 10 -0.222290 x 100 0.262037 X 10-
1
0.604515 x 10
-0.164697 x 10 0.141850 X 10-
1
0.577132 x 10 -0.108842 x 10 0.610581 X 10-'
0.555837 x 10
-0.541356 X 10-
1
0.148800 X 10-'
0.540406 x 10 0.558794 X 10-
8
0.325963 x 10-
A
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
0.283436 x 10
A
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
0.232000 x 10
A
0.232393 x 10
0.232393 x 10.
0.232393 x 10
0.232393 x 10
0.232393 x 10
0.232393 x 100
0.232393 x 100
0.232393 x 10
0.232393 x 10.
0.232393 x 100
0.232393 x 10
APPLICATIONS IN PHYSICS, ENGINEERING, AND BIOLOGY 129
Iteration 4
T X Y MU A
o. 0.100000 X 10
1
-0.640448 X 10 0.232504 X 10 0.232504 X 10
0.1000 0.917827 X 10 -0.558566 X 10 0.175554 X 10 0.232504 X 10
0.2000 0.845336 X 10 -0.482579 X 10 0.129696 X 10 0.232504 X 10
0.3000 0.781763 X 10 -0.411682 X 10 0.933461 X 10-
1
0.232504 X 10
0.4000 0.726438 X 10 -0.345128 X 10 0.648299 X 10-
1
0.232504 X 10
0.5000 0.678777 X 10 -0.288216 X 10 0.428047 X 10-
1
0.232504 X 10
0.6000 0.638277 X 10 -0.222281 X 10 0.262024 X 10-
1
0.232504 X 10
0.7000 0.604510 X 10 -0.164690 X 10 0.141842 X 10-
1
0.232504 X 10
0.8000 0.577121 X 10 -0.108838 X 10 0.610543 X 10-' 0.232504 X 10
0.9000 0.555821 x 10 -0.541333 X 10-
1
0.148790 X 10-' 0.232504 X 10
1.0000 0.540384 X 10 0.335276 X 10-
7
0.791624 X 10-
8
0.232504 X 10
Iteration 5
T X Y MU A
o. 0.100000 X 10
1
-0.640446 X 10 0.232506 X 10 0.232506 X 10
0.1000 0.917826 X 100 -0.558566 X 10 0.175554 X 10 0.232506 X 10
0.2000 0.845336 X 10 -0.482578 X 10 0.129696 X 10 0.232506 X 10
0.3000 0.781763 X 10 -0.411681 X 10' 0.933459 X 10-
1
0.232506 X 10
0.4000 0.726438 X 10 -0.345128 X 10 0.648297 X 10-
1
0.232506 X 10
0.5000 0.678777 X 10 -0.282215 X 10 0.428046 X 10-
1
0.232506 X 10
0.6000 0.638276 X 10 -0.222280 X 10 0.262024 X 10-
1
0.232506 X 10
0.7000 0.604510 X 10 -0.164690 X 10 0.141842 X 10-
1
0.232506 X 10
0.8000 0.577121 X 10 -0.108838 X 10 0.610542 X 10-' 0.232506 X 10
0.9000 0.555820 X 10 -0.541333 X 10-
1
0.148790 X 10-' 0.232506 X 10
1.0000 0.540383 X 10 0.428408 X 10-
7
0.162981 X 10-
7
0.232506 X 10
4. NUMERICAL RESULTS
Taking c = 1, and xo, Yo, flo, and ao as in (3.4), we obtain the
results in Table 6.1. We see that the optimal value of a is given by
(4.1) a '"'-' 0.2325.
(The program is given in Appendix 2.)
To test this conclusion, we set a = 0.2325 in the variational
problem described by (3.1) and (3.2), and solved the resultant varia-
tional problem. This is easily done for the analytic structure we have
carefully chosen, and we obtain the following results, illustrated
graphically, for other nearby values of a. (See Fig. 6.1.) This shows
that the value of a given in (4.1) is indeed optimal.
~
U
0.370
0.360
0.350
0.340 '--_'--_'--_.l...-_.l...-_.l...-_.l....-_..L-_..L----l
o 0.10 0.20 0.30 0.40
System design parameter, a
Fig. 6. I-Cost as a function of design parameter
5. AN INVERSE PROBLEM IN RADIATIVE TRANSFER
Let us now consider a particular version of what are called
"inverse problems." In the two previous volumes in this series,
we discussed the problem of determining the reflected flux from
an inhomogeneous, plane-parallel, nonemitting, and isotropically
scattering atmosphere of finite optical thickness Tb where the optical
properties depend only upon T, the optical height. (See Fig. 6.2.)
Let parallel rays oflight of net flux 7T per unit area normal to their
direction of propagation be incident on the upper surface in the
Incident Flux Reflected Flux
\\ \\/1
" {
} ,
Fla. 6.1-The physical situation
APPLICATIONS IN PHYSICS, ENGINEERING, AND BIOLOGY
131
direction characterized by flo, where flo is, as usual, the cosine of the
angle measured from the inward normal to the surface. The bottom
surface, for simplicity, is taken to be completely absorbing.
Formulating the process in invariant imbedding terms, let r(fl,flo,Tl)
denote the intensity of the diffusely reflected light in the direction
cos-
1
fl and set R(fl,flo,Tl) = 411'. Then R satisfies the integro-
(5.1)
oR ( -1 -1)
- = - fl + flo R
OTI
[
1 (1 d 'J
+ A(Tl) 1 + '2 Jo R(fl,fl',T1) ;,
[
1 (1, dfl'J
. 1 + '2 Jo R(fl ,flo,Tl);: ,
with the initial condition R(fl,flo,O) = O. The function A(T) is the
albedo for single scattering.
To obtain a numerical solution of (5.1) via digital computers, we
apply Gauss quadrature of order N to the integrals, and treat the
resulting system of ordinary differential equations
(5.2) dd Rij(Tl) = -(fli
1
+ fljl)Rij
Tl
+ A(T
1
) [1 + ! !R;kh) WkJ
2 k=1 flk
Rij(O) = 0,
where
(5.3)
Suppose, on the other hand, that we are given the values RitCTl),
i,j = I, 2, ... , N, and are asked to determine the function A(T),
o :5: T :5: T
1
This is the type of question that arises naturally in
astrophysics in connection with various types of experimental work.
6. ANALYTIC FORMULATION
In order to tackle a problem of this nature, we must impose some
analytical structure on A(T), which is to say, we must add some further
information concerning the nature of the physical process. To
illustrate, consider the case where the medium consists essentially of
two layers, each with constant albedo, separated by a thin zone of
rapid transition from one value of the albedo to the other. * Let the
albedo have the form .
(6.1) A(T) = a + b tanh [!O(T - c)],
where a, b, and c are constants to be determined. We see that
Al ~ a - b in Layer I, and A2 ~ a + b in Layer 2.
Given the N2 values bij = r;;(T
I
), we would like to determine the
three parameters a, b, and c, and Tb the thickness. To this end, we
ask for the values of a, b, c, and TI which minimize the mean-square
error
(6.2) S = L {rilTI) - b
i
;}2,
i,;
where Ri;(TI) = 4,uir;;(TI), and the Ri;(TI) satisfy the equation of (5.2).
This variational problem is treated below by means of the quasi-
linearization techniques we have previously discussed.
We begin by generating "observations" by choosing a = 0.5,
b = 0.1, c = 0.5, and integrating to a thickness of TI = 1.0, using
N = 7. We now carry out three types of numerical experiments:
(a) Given a, b, and Tl, determine c, the altitude of the interface.
(b) Determine Tb the over-all thickness, given a, b, and c.
(c) Determine a, b, and c, given TI'
The importance of these problems was pointed out to the authors by
S. Ueno of Kyoto University, Japan.
APPLICATIONS IN PHYSICS, ENGINEERING. AND BIOLOGY
133
As will be seen from the following brief tabulation of the results
for cases (a) and (c) in Tables 6.2 and 6.3, the quasilinearization
technique works well. The success of the method depends crucially
upon the initial approximation. In other words, we can only expect
Table 6.1
SUCCESSIVE APPROXIMATIONS OF c, THE LEVEL
OF THE INTERFACE
Approximation Run 1 Run 2 Run 3
0 0.2 0.8 0.0
1 0.62 0.57
2 0.5187 0.5024 No
3 0.500089 0.499970 convergence
4 0.499990 0.499991
True Value 0.5 0.5 0.5
Table 6.3
SUCCESSIVE APPROXIMATIONS OF AI' A
2
AND c
Approximation Al=a-b A
2
=a+b c
0 0.51 0.69 0.4
1 0.4200 0.6052 0.5038
2 0.399929 0.599995 0.499602
3 0.399938 0.599994 0.499878
True Value 0.4 0.6 0.5
the method to be successful if initially we have at least a crude idea
as to the nature of the actual physical process. (The programs are
given in Appendix 3.)
The quasilinearization technique in case (c), with N = 7, involved
the integration of 124 linear differential equations, and the solution
of a system of linear algebraic equations of order 3. The calculations
took about two minutes on an IBM-7044.
Subsequently, we shall indicate how a combination of dynamic
programming and quasilinearization will lead to an improved tech-
nique for determining the positions ofthe interfaces in a multilayered
medium.
8. THE VAN DER POL EQUATION
One of the most famous of all nonlinear differential equations is
the Van der Pol equation,
(8.1) x" + A(X
2
- l)x' + x = o.
Although a nonlinear equation equivalent to this had been studied
by Rayleigh, this particular equation achieved prominence only in
connection with the development of modern electronics. The equation
provides an excellent description of the behavior of a multivibrator,
and thus has application to many other scientific fields, such as
cardiology or electroencephalography, in which similar "flip-flop"
phenomena are observed. This was pointed out by Van der Pol
himself in his original paper.
It is of some interest then to ask whether or not observation of the
solution can yield an effective determination of the parameter A.
Assume that the following three observations of the displacement x
have been made:
(8.2) x(4) = -1.80843,
x(6) = -1.63385,
x(8) = -1.40456.
We wish to determine both A and u(4) = x'(4). It is not necessary to
use least-squares, since only three observations are given. To obtain
an initial approximation, we observe that
(8.3) x(6) ~ x(4) ~ 0 . 0 8 7 ~ x'(4),
x(8) - x(6) ~ 0.114 ~ x'(6),
2 - -
x'(6) - x'(4) ~ 0.014 ~ x"(4).
2 - -
APPLICATIONS IN PHYSICS, ENGINEERING, AND BIOLOGY 135
Hence, referring to (8.1), we obtain
(8.4) Ao = 7,
as an initial guess for A. We next integrate (8.1), or the equivalent
system
(8.5) x' = u,
u' = _A(X2 - 1)x' - x,
A' = 0,
subject to the initial conditions
(8.6) x( 4) = -1.80843,
u(4) = +0.08,
A(4) = 7.0.
To obtain the (n + l)st approximation, after having calculated the
nth, we use the linearized relations
(8.7) xn+! = un+!'
U
n
+
1
= -An(X! - 1)u
n
- Xn + (Xn+l - xn)[-2AnXnUn - 1]
+ (Un+! - Un)[-An(X! - 1)]
+ (An+l - An)[-(X! - 1)u
n
],
in+l = 0,
together with the three-point boundary conditions of (8.2).
The results of a numerical experiment are summarized in Table 6.4.
A second experiment was carried out with poorer estimates of the
initial velocity and system parameter. The results are presented in
Table 6.5. A third trial in which the initial estimate of the system
parameter was taken to be 20 resulted in an overflow, so that no
results are available.
The data in Eq. (8.6) were generated by integrating Van der Pol's
equation with
(8.8) A = 10.0,
x(O) = 1.0,
X(O) = 0.0.
A graph is presented in Fig. 6.3. (The programs are given in
Appendix 4.)
Table 6.4
THE FIRST NUMERICAL EXPERIMENT
Initial Iteration Iteration Iteration True
Approx. 2 3 4 Values
x(4) -1.80843 -1.80843 -1.80843 -1.80843 -1.8084322
u(4) +0.08 +0.0564454 +0.0794911 +0.079366 0.079366909
).(4)
+7.0 9.91541 10.0004 10.00000 10.00000
Table 6.5
THE SECOND NUMERICAL EXPERIMENT
Initial Iteration Iteration Iteration True
Approx. 2 4 6 Values
x(4) -1.80843 -1.80843 -1.80843 -1.80843 -1.8084322
u(4) 0.1000 -2.0758 0.599288 0.0792063 0.079366909
),(4) 5.000 3.87992 11.4091 9.99956 10.000
2.0
1.0
i
o r - ~ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ~ - - - - - - - - - - - - - 4
-1.0
-2.0
o 2 4 6 8 10 12 14
Fig. 6.3-Solutlon of the Van der Pol equation for). = 10.0, x(O) = 1.0,
X(O) = 0.0
9. ORBIT DETERMINATION AS A MULTIPOINT
BOUNDARY-VALUE PROBLEM
The general problem of solving the vector equation
(9.1) x' = f(x,t),
subject to the multipoint conditions
(9.2)
N
Laijxlt
i
) = bi'
;=1
o ~ t ~ b,
j = 1,2, ... , N,
137
whereO ~ tl ~ t2 ~ .. ~ tN ~ b, has been mentioned previously in
connection with quasilinearization. Let us now consider a particular
version of some interest occurring in orbit determination. Suppose
that the equation of motion of a heavenly body is, in normalized units,
(9.3) x' = u, , y' = v,
u' - _ x v' = _ y
- (x2 + y2)1.5 ' (x2 + y2)1.5
The motion takes place in the (x,y)-plane, the plane of the ecliptic.
At times ti' i = 1,2,3,4, we are given the quantities (J(t
i
) = (J;, which
yield the conditions
(9.4) yeti) = (x(t
i
) - 1) tan (Ji' j = 1,2,3,4.
(See Fig. 6.4.)
y
H (x, y)
------------.. - - - - - - ~ __ - - ~ - - - - - - - - - - - - ~ x
Sun
Fig. 6.4-The physical situation
We wish to determine the orbit in convenient form by obtaining
the values of x, y, U, v at time t = 2.5. For simplicity, we are assuming
that the earth is stationary at (1,0), and that the sole force on the
heavenly body is the gravitational attraction of the sun.
Consider the case for which the observational data are (in radians)
(10.1) 6(0.5) = 0.251297,
6(1.5) = 0.783690,
6(1.0) = 0.510240,
6(2.0) = 1.076540.
These data were generated by assuming that at time zero we have
x(O) = 2.0, yeO) = 0, u(O) = 0, v(O) = 0.5. The situation, with a
stationary earth, is as shown in Fig. 6.5. As an initial approximation
to the trajectory, we assume that the observed body is at the point
(1.0,0) at time zero and that it moves in a circular orbit with a period
equal to 3.664, a very poor initial approximation. Then, using a
Sun
Earth
Fig. 6.S-Approxlmate directions to the heavenly body at four times
of observation from a fixed earth
139
Runge-Kutta integration method and the quasilinearization pro-
cedure, we found the results shown in Table 6.6.
The calculation took about one and one-half minutes, * no attempt
having been made to streamline the calculations. It produced, of
course, not merely the results of Table 6.6 but an ephemeris for
o ~ t ~ 2.5 at intervals of 0.01. (The programs are given in Appendix
5.)
x(2.5)
u(2.5)
y(2.5)
v(2.5)
Table 6.6
THE APPROXIMATIONS TO THE DISPLACEMENTS AND
VELOCITIES AT TIME 2.5
Initial First Second Third
Fourth
Precise
Value to
Approx. Approx. Approx.
Approx.
Approx.
6 Figures
-0.455175 1.554360 1.169660 1.190170 1.193560 1.193610
-0.873320 -1.638830 -0.495101 -0.656686 -0.664207 -0.664263
0.975790 2.982690 0.865184 1.043930 1.060540 1.060700
-0.408690 1.807620 0.201206 0.237695 0.247433 0.247499
Perhaps even more surpnsmg is the fact that if, as an initial
approximation to the orbit, we consider the heavenly body to
coincide with the position of the earth over all time, we still obtain
the same rapid convergence to the correct orbit.
The method is readily extended to cases where more than four
observations are available through use of the method ofleast-squares.
Furthermore, perturbations due to the motion of the earth and the
presence of other planets are readily incorporated, since they merely
result in modifications of equations (9.3).
II. PERIODOGRAM ANALYSIS
A basic problem in the scientific world is that of detecting hidden
periodicities. Indeed, it is difficult to think of a domain where
questions of this nature do not arise in important ways. Analytically,
the problem takes a very simple form. Given the information that a
On an IBM 7090.
function J(t) has the structure
R
(11.1) J(t) = cos wit,
;=1
we wish to determine the amplitudes and frequencies, and R, from
a set of observations, b
i
= J(ti), i = 1,2, ... , M. We shall consider
the simpler problem where R is assumed known and take M 2R.
Our approach, which is different from any of the standard ones,
is to convert this into a variational problem involving multipoint
boundary values and then to employ quasilinearization. We begin
by writing
R
(11.2) J(t) = !utC
t
),
i=l
where the U
i
are solutions of
(11.3) = o.
Our aim is to determine the and the Wi so as to minimize the
quantity
M
(11.4) ! (J(t
i
) - b
i
)2.
i=l
To treat this, we employ quasilinearization as in the preceding pages,
starting with initial approximations which produce the
initial functions U
i
(O)(t). In the next section, we present some numerical
results.
Let us consider a simple case where
(12.1)
with
(12.2)
3
J(t) =
i=1
= 1.000000,
= 0.500000,
(Xa = 0.100000,
WI = 1.11000,
W2 = 2.03000,
Wa = 3.42000,
and six observations,
t
0.00
1.00
2.00
3.00
4.00
5.00
f(t)
l.6oooo
0.126895
-0.823201
-0.558708
-0.356338
0.351082
14
The results of the successive approximations are given in Table 6.7
Table 6.7
Approximation i (Xi
Wi
0 1 0.900000 1.00000
2 0.600000 2.00000
3 0.200000 3.00000
1 1 0.943582 1.09000
2 0.486676 1.92583
3 0.169742 2.45651
--
2 1 1.00218 1.11113
2 0.476529 2.01352
3 0.121287 2.62338
3 1 1.00002 1.11001
2 0.503119 2.03179
3 0.096864 2.80801
--
4 1 0.999988 1.11000
2 0.499537 2.02981
3 0.100475 2.85931
True 1 1.000000 1.11000
2 0.500000 2.03000
3 0.100000 3.42000
The case of seven observations,
t
0.00
0.83
1.67
2.50
3.33
4.17
5.00
yielded the results shown in Table 6.8.
Table 6.8
Approximation i (Xi
0 1 0.900000
2 0.600000
3 0.200000
1 1 0.984727
2 0.535234
3 0.057545
2 1 0.998859
2 0.491950
3 0.107930
3 1 1.00904
2 0.500095
3 0.099857
4 1 1.000000
2 0.500000
3 0.100000
True 1 1.000000
2 0.500000
3 0.100000
J(t)
1.60000
0.452413
-0.679691
-0.820313
-0.367504
-0.381874
0.351082
Wi
1.00000
2.00000
3.00000
1.14378
2.07572
3.29235
1.11124
2.02810
3.39973
1.11003
2.03018
3.41930
1.11000
2.03000
3.42000
1.11000
2.03000
3.42000
The case of eight observations,
t
0.00
0.71
1.43
2.14
2.86
3.57
4.29
5.00
produced the results given in Table 6.9.
Table 6.9
Approximation i (Xi
0 1 0.900000
2 0.600000
3 0.200000
1 1 0.981747
2 0.531127
3 0.056783
2 1 0.998520
2 0.492646
3 0.105177
3 1 1.00006
2 0.500003
3 0.099888
4 1 1.000000
2 0.500000
3 0.100000
True 1 1.000000
2 0.500000
3 0.100000
f(t)
1.60000
0.694147
-0.484599
-0.849519
-0.649072
-0.302561
-0.378655
0.351082
Wi
1.00000
2.00000
3.00000
1.13890
2.06075
3.22805
1.11046
2.02508
3.41224
1.11003
2.03016
3.41981
1.11000
2.03000
3.42000
1.11000
2.03000
3.42000
14
13. DISCUSSION
It was found that a reasonably good initial approximation is
required to obtain convergence. For example, when experiments
similar to those described in Tables 6.7 and 6.8 were carried out with
the initial values
(13.1)
it was found that the successive approximations did not converge.
14. PERIODIC FORCING TERMS
Similar techniques can be used to determine the !Xi and Wi in the
forcing term, when the function under observation, u(t) , is given as
a solution of a linear differential equation of the form
(14.1) u" + blu' + b
2
u = !Xl cos WIt + !X2 cos w
2
t,
u(O) = Cl> u'(O) = C
2

Actually, we can do very much better. Given a set of observations,
{u(t
i
)}, we can determine the initial conditions C
h
C
2
, the system
parameters bI> b
2
, and the parameters in the forcing function.
15. ESTIMATION OF HEART PARAMETERS
Recently it has been found possible to account for the potentials
which are observed on the skin of a patient and which are due to
ventricular depolarization. The basic idea is to divide the ventricles
into a number of segments, each of which contains a dipole source
of current. The dipole is located in the center of the segment, and its
axis is orthogonal to the surface of its segment. The dipoles have
time-varying moments which are dependent upon the depolarization
wave which sweeps over the ventricles. These concepts are illustrated
in Figs. 6.6 and 6.7. The postulated sequence of depolarization events
RIGHT VENTRICLE
Time sequence
human
myocardial
depolarization
o
SEPTUM LEFT VENTRICLE
R. V. apex
,. --I
L. V. apex
,- --I
Body of R. V.
I- __ I
,.. Body of L. V. ....,
Septum - L-R
Millisec
Fig. 6.6-Myocardial segments
145
is obtained via an extrapolation of the well-known experimental
results of A. Scher for canine hearts. The body is viewed as a volume
conductor. Surface potential time histories can be produced, in this
fashion, which are in good agreement with those obtained from both
normal and abnormal hearts, as has been demonstrated by numerous
high-speed computer runs.
Current
Dipole
Moment
~ - - - - - - - - - - - - - - - - a - - - - - - - - - - - - - - - - - - - - - ~ T i m e
Fl 6.7-Typlcal dipole moment function F;(t)
The aim of the present discussion is to show that it is also possible
to solve the inverse problem; that is, given the time histories of the
surface potentials we can deduce the time histories of the individual
dipole moments. From the mathematical viewpoint we treat this as a
nonlinear multipoint boundary-value problem. It is resolved using
quasilinearization following the lines laid down in the previous pages.
16. BASIC ASSUMPTIONS
Introduce standard orthogonal reference axes, x, y, z, and consider
the dipole located in the ith segment at point (Xi,Y;,Zi) and having
(li,m;,n
i
) as the direction cosines of its axis. This produces a potential
w at the point (aj,bj,c
j
) which depends upon the assumptions made
about the body and its surrounding medium as volume conductors.
If, for simplicity, we consider the body to be homogeneous and of
infinite extent, then we may write
(16.1)
where
(16.2) (a) k = const,
(b) Fi(t) = moment of dipole in ith segment at time t,
(d) 'i
2
j = (ai - Xi)2 + (b
j
- Yi)2 + (e
j
- Zi)2.
Points infinitely far from the dipoles are at zero potential. Let us
introduce a reference point (ao,bo,c
o
) and assume that at N surface
points the potentials with Iespect to that at the reference point are
determined. In addition, we assume that the ventricles are divided into
M segments.
Now we must consider the dipole moment functions Fi(t) , i =
1, 2, ... , M, in more detail. One is pictured in Fig. 6.7. We have
found that these curves are well represented as solutions of the
147
ordinary differential equations
(16.3) o ~ t ~ 80 msec,
i = 1,2, ... , M.
The parameters ti are the times at which the maxima of the moments
occur, the k
i
are related to the broadness of the curves, and the hi
are related to the maxima of the moments. The analytical solution
of these equations is of no interest to us, since in more refined studies,
(16.3) will be replaced by much more complex differential equations.
We denote the potential produced at time t at the jth observation
point by Vlt),j = 1,2, ... ,N, with 0 ~ t ~ 80 msec. We have
M M
(16.4) Vlt) = k ! (Flt) cos ()ii)rij2 - k ! (Fi(t) cos ()i)r;2,
;=1 i=1
j = 1,2, ... , N,
where
(16.5) (a) cos ()i = [li(aO - Xi) + m;(b
o
- Yi) + n;(c
o
- zi)]r;-t,
(b) r ~ = (a
o
- Xi)2 + (b
o
- Yi)2 + (co - Zi)2,
i = 1,2, ... , M.
17. THE INVERSE PROBLEM
Assume that a heart has been observed for 0 ~ t ~ 80 msec and
that the surface potential time series are recorded; that is,
bj(t) = observed surface potential at time t
at the jth observation position,
j = 1, 2, ... , Nand t = 0, 1, 2, ... , 80 msec.
We wish to determine the current dipole moment functions Fi(t),
i = 1, 2, ... , M, which produce surface potentials that are in best
agreement with the observed potentials. Let us put this in more
precise mathematical language. We wish to determine the 3M
constants ki' Ii' and hi in the differential equations of (16.3) which
are such that the sum of the squares ofthe differences ofthe potentials
produced at the observation points and the observed potentials is as
small as possible. If
80 N
(17.1)
S =! !(V;(r) - birW,
r=O ;=1
where the functions Vj(t) are determined from (16.4), and Fi(t) are
determined from (16.3), we wish to minimize S through an appropri-
ate choice of the heart parameters't
i
, ki' and hi, i = 1,2, ... , M.
18. NUMERICAL EXPERIMENTS
Earlier computational work led us to believe that we could handle
the inverse problem in the manner described in previous sections.
Some numerical experiments were conducted to verify this. First, we
chose M, the number of ventricular segments, equal to five and N,
the number of observation points, equal to three. Then various values
were assigned to the parameters ti' ki' and hi, i = 1, 2, ... , 5, to
simulate a normal heart and abnormal hearts exhibiting hypertrophies
and infractions. Also values were assigned to (Xi,Yi,Zi), (a;,bj,c
j
), and
(/i,mi,ni)' By integrating (16.3) with these values, and using (16.4)
and (16.5), we produced sets of hypothetical surface potentials
corresponding to a variety of hearts. We used segments 1, 8, 11, 14,
and 20, and these five segments of the heart produced good simulation
of potentials measured by three orthogonal leads in normal human
subjects. (See Fig. 6.8.)
We assumed that the values t
i
, i = 1,2, ... , 5, are known, which
we take to represent a heart with a normal conduction system. We
also assumed a set of surface potential measurements b;(r),} = 1,2,
3; r = 0, 1,2, ... , 80. Then use of the method of quasilinearization
did produce the "missing" values of the heart parameters, k
i
and hi,
i = 1,2, ... , 5. It should be borne in mind that the computational
load is considerable. We are required to solve 110 simultaneous
linear differential equations and to solve linear algebraic equations
'"
-
c:
:J
g>., O ~ - - P ~ - - - - - - - - - - : : : 0 0 0 1
~
g
o 20 40 60 80
t, millisec
149
Fig. 6.S-Potentials produced by five-segment normal heart model
of order ten at each stage in the calculation. Much remains to be
done to investigate the range of convergence of the method, to deter-
mine the effects of errors in the observed surface potentials, and to
make the computational program as efficient as possible. These pre-
liminary experiments merely indicate feasibility of this approach to
the formulation and numerical solution of inverse problems in cardi-
ology. (Some programs are given in Appendix 6.)
Chapter Six
2. See
R. BELLMAN, H. KAGIWADA, and R. KALABA, "A Computational Procedure for
Optimal System Design and Utilization," Proc. Nat. Acad. Sci. USA, Vol. 48,
1962, pp. 1524--1528.
5. This material was first presented in
R. BELLMAN, H. KAGIWADA, R. KALABA, and S. UENO, Inverse Problems in
Radiative Transfer: Layered Media, The RAND Corporation, RM-4281-
ARPA, December 1964 (to appear in Icarus).
R. BELLMAN, H. KAGIWADA, R. KALABA, and S. UENO, On the Identification of
Systems and the Unscrambling of Data-II: An Inverse Problem in Radiative
Transfer, The RAND Corporation, RM-4332-ARPA, November 1964.
The values of the reflection matrix for N = 7 and the albedo function 11.( r) =
0.5 + 0.1 tanh 100r - 0.5) for a slab of thickness rl = 1.0 are given below in
the table, "The Measurements {b
ii
}."
THE MEASUREMENTS {hi;}
j=1 2 3 4 5 6 7
i = 1 0.079914 0.028164 0.014304 0.009104 0.006707 0.005515 0.004970
2 0.143038 0.091522 0.058437 0.040826 0.031405 0.026378
0.023989
3 0.167000 0.134331 0.099653 0.075106 0.060044 0.051445 0.047248
4 0.178898 0.157955 0.126408 0.099392 0.081253 0.070435
0.065042
5 0.185284 0.170817 0.142072 0.114229 0.094495 0.082423
0.076332
6 0.188723 0.177733 0.150791 0.122665 0.102104 0.089349
0.082870
7 0.190354 0.180898 0.154995 0.126773 0.105829 0.092748
0.086083
It is interesting to compare them with the corresponding values tabulated in
Vol. I of this series for homogeneous slabs of thickness 1.0 and albedos 0.4, 05.,
and 0.6. There are significant differences introduced by the inhomogeneities.
8. Our discussion is based on
R. BELLMAN, H. KAGIWADA, and R. KALABA, Quasilinearization, System Identi-
fication, and Prediction, The RAND Corporation, RM-3812-PR, August 1963.
An interesting application of these techniques is to the field of electro-
encephalography. For the formulation, see
E. M. DEWAN, "Nonlinear Oscillations and Electroencepholography," J.
Theoret. Bioi., Vol. 7, 1964, pp. 141-159.
151
9. This approach was sketched in
R. BELLMAN, H. KAGIWADA, and R. KALABA, "Orbit Determination as a
Multipoint Boundary-value Problem and Quasilinearization," Proc. Nat.
A cad. Sci. USA, Vol. 48, 1962, pp. 1327-1329.
An interesting programming advance in the determining of the partial derivatives
is given in
R. BELLMAN, H. KAGIWADA, and R. KALABA, Wengert's Numerical Method for
Partial Derivatives, Orbit Determination, and Quasilinearization, The RAND
Corporation, RM-4354-PR, November 1964 (to appear in Comm. ACM).
For another application, see
A. LAVI and J. C. STRAUSS, Parameter Identification in Continuous Systems, to
appear.
11. This section is based on
R. BELLMAN, H. KAGIWADA, and R. KALABA, On the Identification of Systems
and the Unscrambling of Data-I: Hidden Periodicities, The RAND Corpora-
tion, RM-4285-PR, September 1964.
15. The mathematical model is presented in
R. SELVESTER, C. COLLIER, and R. PETERSON, "Analogue Computer Model of the
Vectorcardiogram," Circulation Research, to appear.
This section is based on joint research reported in
R. BELLMAN, C. COLLIER, H. KAGIWADA, R. KALABA, and R. SELVESTER,
"Estimation of Heart Parameters Using Skin Potential Measurements,"
Comm. ACM, Vol. 7, 1964, pp. 666-668.
We avoided any discussion of recent applications of a combination of dynamic
programming and quasilinearization to system identification. See
R. BELLMAN, B. Gwss, and R. ROTH, Identification of Differential Systems with
Time-varying Coefficients, The RAND Corporation, RM-4288-PR, November
1964.
R. BELLMAN, B. Gwss, and R. ROTH, "Segmental Differential Approximation
and the 'Black Box' Problem," J. Math. Anal. Appl., to appear.
R. BELLMAN, B. GLUSS, and R. ROTH, "On the Identification of Systems and the
Unscrambling of Data: Some Problems Suggested by Neurophysiology,"
For the application of Laplace transform methods, see
R. BELLMAN, H. KAGIWADA, and R. KALABA, "Identification of Linear Systems
Using Numerical Inversion of Laplace Transforms," IEEE Trans., to appear.
Chapter Seven
DYNAMIC PROGRAMMING AND
QUASILINEARIZATION
I. INTRODUCTION
In this chapter, we wish to discuss some of the connections
between dynamic programming and the theories of positive operators
and quasilinearization. To begin with, we trace the origin of quasi-
linearization in the concept of approximation in policy space.
Following this, we will indicate how a combination of quasilineariza-
tion and dynamic programming can be used to provide an approxi-
mation method for the solution of multidimensional variational
problems which greatly increases the scope of the individual theories.
Furthermore, we show how to combine dynamic programming and
quasilinearization to treat some complicated identification problems.
References to detailed discussions will, as in previous chapters, be
found at the end of the chapter. .
2. A BASIC FUNCTIONAL EQUATION
Let us consider a multistage decision process taking place in a
state space S, where pES represents a typical state, T(p,q) is an
element of a family of transformations on S characterized by q, the
policy vector, qED, the policy or decision space, and g(p,q) rep-
resents the single-stage return.
If we denote by!(p) the total return from an unbounded process,
obtained using an optimal policy, the principle of optimality yields
the functional equation
(2.1) f(p) = max [g(p,q) + f(T(p,q))].
Q
153
A question of immediate interest is that of determining the existence
and uniqueness of solutions of this equation under various as-
sumptions concerning the functions g and T and the spaces Sand D.
A variety of methods in classical analysis can be called upon for
this purpose, such as fixed-point theorems, approximation by discrete
processes, and the method of successive approximations. Thus, for
example, using the last method named, we can study (2.1) by means
of the sequence {fN(P)} generated by
(2.2) fN(P) = max [g(p,q) + fN-l(T(p,q))], N ~ 1,
q
with !o(p) chosen in some adroit fashion.
This is a meaningful approximation technique because it IS
equivalent to considering the unbounded process as a limit of
bounded processes.
3. APPROXIMATION IN POLICY SPACE
One of the advantages possessed by the theory of dynamic pro-
gramming is that it permits, and indeed encourages, a type of approxi-
mation, called approximation in policy space, which does not exist
in classical analysis. The basic idea is quite simple. Instead of
approximating to the function !(p), we approximate to the policy
function q(P). From the standpoint of applications, this is a very
natural approach. As we shall see in a moment, it possesses an
important analytic advantage.
Let us guess an initial policy qo(P) and determine the corresponding
return function!o(p) by means of the functional equation
(3.1)
To obtain an improvement of qo(P), we ask for the function ql(P)
which maximizes the expression
(3.2) g(p,q) + fo(T(p,q)).
DYNAMIC PROGRAMMING AND QUASILINEARIZATION
ISS
Using this policy, we generate the functionjl(p) as the solution of
(3.3)
As in (3.1), the solution is generated by straightforward iteration,
In this fashion, we generate a sequence of policies {qN(P)} and return
functions {fN(P)}.
Let us now examine what is required to justify the statement that
ql(P) is an improvement on qo(p). We have
(3.5) fo(p) = g(p,qo) + fo(T(p,qo))
~ g(P,ql) + fo(T(p,Ql))'
by virtue of the way in which ql was chosen. Comparing this with
the equation of (3.4) for !t(p), we see that the inequality /0 ~ j l is
now a consequence of a positivity relation for the operator
(3.6) T(f) = f(p) - f(T(p,q)).
In this way, then, approximation in policy space leads us to the area
of functional inequalities. Once having seen the advantages that
accrue, it is a simple step to the procedures described in the previous
chapters for converting descriptive processes into decision processes.
Assuming that T(j) satisfies the requisite positivity properties, let
us show that each functionjip) is bounded by j(p). We have
(3.7) fn(P) = g(p,qn) + fn(T(p,qn)),
f(p) = max [g(p,q) + f(T(p,q))]
q
from which the desired inequality follows.
This monotonicity of convergence furnishes us with a very flexible
way of establishing the existence of solutions of (2.1).
4. DYNAMIC PROGRAMMING AND QUASILINEARIZATION
The principal difficulty in the application of the theory of quasi-
linearization lies in the restrictive conditions required to ensure con-
vergence of the sequence of successive approximations. The examples
given in Chapter Six, however, illustrate the fact that these methods
often work better in practice than might be expected. Nevertheless,
this is not a very satisfactory state of affairs and it is essential to
develop foolproof (or should we say divergence-proof) techniques.
Suppose, for instance, that we wish to apply quasilinearization to
the minimization of the functional
(4.1) J(u) = LT g(u,u') dt,
where u is a scalar function constrained by the terminal conditions
u(O) = Co, u(T) = c'. Ifwe take uo(t), the initial approximation, to be
the straight line joining (O,co) and (T,c'), then to guarantee con-
vergence we must put a bound on T. If the initial approximation
were more adroitly chosen, we could afford a larger value of T.
Unfortunately, in many applications, we do not possess sufficient
information to do better initially than some simple choice such as
the foregoing.
To overcome this difficulty, let us use a blend of quasilinearization
and dynamic programming. Essentially, we intend to apply quasi-
linearization for local optimization, and dynamic programming to
obtain global optimization.
To illustrate the combination of techniques, let us return to the
variational problem posed above. In the (c,T)-plane, we construct
a grid of points as indicated in Fig. 7.1.
We have kept the number of points small (and the example one-
dimensional) in order to present the ideas clearly without complicat-
ing the diagrams or notation. In practice, we could very easily
accommodate grids of several thousand points in phase space in the
treatment of multidimensional variational problems.
DYNAMIC PROGRAMMING AND QUASI LINEARIZATION
157
C
Cl C, C}U
C,
C" CII
Cn
C'
CII CII
o 2S 3S 4S
T=5S
Fig. 7.1
Let us suppose that from coone can go only to one of three points
Cl>C
2
,C
3
; from C
b
C
2
,C
3
to one of the points C
4
'C
S
'C
6
, and so on; from
C1o,C
ll
, or C
12
, one must go to c'. The allowable paths from one grid
point to another are those that minimize the functional
(4.2)
l
(k+1)6
J(u) = g(u,u') dt,
k6
with u(kb) = C
i
, u(k + l)b = c;, with C
i
and C; "nearest neighbors"
in the foregoing sense. Call these paths geodesics and the correspond-
ing value of J(u) a geodesic length. Let d
i
; denote this length.
We have previously indicated how the geodesic distances can be
determined via quasilinearization if b is sufficiently small. The bound
on b can be determined a priori in terms of bounds on g and its
partial derivatives. Let us now show how to use dynamic programm-
ing to piece these geodesic arcs together to find an excellent approxi-
mation to the actual minimizing function.
5. FUNCTIONAL EQUATIONS
The optimization problem posed in the preceding section is a
special case of a more general routing problem for which very
effective algorithms exist. In this case, the structure of the mesh
makes the calculation particularly simple.
Introduce the function
(5.1) f. = the minimum total geodesic length from C
i
to c',
defined for i = 0, 1, ... , 12. Then the principle of optimality yields
the functional equation
(5.2)
h = min [dij + Ii]'
i*i
connecting the f.. In this case, this equation reduces to a recurrence
relation, since f. is immediately determined for the points (c
10
,C
U
'C
12
).
For these, we have
(5.3)
Using these values in (5.2), we can calculate f. for (c
7
,C
S
,C
9
). Con
tinuing in this fashion, we calculate (c
4
,C
S
,C
6
), (C
1
,C
2
,C
3
), and finally
Co, the desired value.
6. DISCUSSION
Once the quantities dij have been determined, the determination
of the f. requires negligible time. The time required to obtain the
d
ij
depends directly, of course, upon their number, which, in turn,
depends on Cl. There are, however, many techniques that can be used
to reduce the time of computation.
Having determined the optimal "polygonal" path in this fashion,
we can, if we wish, use it as the initial approximation in an attack
on the original problem via quasilinearization. In many cases, the
"polygonal" path will provide a sufficiently accurate solution.
7. DYNAMIC PROGRAMMING AND DIFFERENTIAL
APPROXIMATION
Let us now rapidly sketch ways in which we can combine dynamic
programming and differential approximation. Suppose that the
159
problem we are considering is that of determining the parameters a
and c so that the solution of
(7.1)
dv
- = g(v,a),
dt
v(O) = C,
is close to the given function u(t) in the sense of minimizing the
expression
(7.2)
As in the minimization problem discussed in the previous sections,
the quasilinearization technique can fail if the initial approximation
is not sufficiently close.
One reason for this failure may be that u(t) is really composed of
parts of a number of distinct processes. Thus, if we write vet) =
v(t,a,c) (denoting explicitly the dependence on a and c), it may well
be that
(7.3) o ~ t ~ t
1
,
t1 ~ t ~ t
2
,
= v(t,aN,cN), t
N
-
1
~ t ~ tN.
Knowing that u(t) has this structure, a quite typical situation in
many cases, how do we determine the quantities ai' c
i
, and ti? This
is a system identification problem.
8. DYNAMIC PROGRAMMING AND THE IDENTIFICATION
OF SYSTEMS
To indicate the way in which we intend to apply dynamic pro-
gramming, let us consider perhaps the simplest version of the forego-
ing problem, that of obtaining optimal polygonal approximation to
a given function u(t) over an interval [0, T]. Let us suppose that we
allow a polygonal curve with (N + I) line segments, as indicated in
u
.... _-
T
o t,
Fig. 7.1
Fig. 7.2. We wish to determine the N values, t
i
, and the 2N param-
eters, a
i
and hi, so as to minimize the quantity
(8.1)
N Jti+l
DN = tj (u(t) - a
i
- b
i
t)2 dt.
Here to = 0, t
N
+
l
= T. Write
(8.2)
for any two numbers Sl, S2 satisfying 0 :::;; Sl :::;; S2 :::;; T. The function
is readily obtained by means of the two linear algebraic
equations obtained by taking partial derivatives with respect to a
i
and hi'
Introduce the functions
(8.3) fN(a) = min D
N
,
ai,bi,ti
defined for N = 0, 1, 2, ... , a O. Then the principle of optimality
readily yields the recurrence relation
(8.4) N 1,
fo(a) =
Once the functions have been calculated, the determination
of the sequence {!N(a)} is quite routine.
161
9. DISCUSSION
The quantity ~ ( S l h ) arises in connection with the approximation
of u(t) by means of a solution of
d
2
v
(9.1) dt
2
= 0, v(o) = a, v'(O) = b.
This is a very simple form of differential approximation. If we wish
instead to use the solution of (8.1) for curve fitting, which is to say,
to engage in more sophisticated differential approximation, we re-
quire the services of quasilinearization as described in Chapter Six.
Once again, there are a number of techniques we can use to decrease
the time required to compute ~ ( S h S 2 ) for a grid of (slh)-values in
o ~ Sl ~ S2 ~ T.
Let us note finally that we could use a Cebycev criterion,
(9.2) min max lu(t) - a
i
- bitl
in place of the mean-square criterion without in any way increasing
the computational burden.
Chapter Seven
1-3. For an introduction to the theory of dynamic programming, with
applications to numerous areas, see
R. BELLMAN, Dynamic Programming, Princeton University Press, Princeton,
New Jersey, 1957.
R. BELLMAN and S. DREYFUS, Applied Dynamic Programming, Princeton Univer-
sity Press, Princeton, New Jersey, 1962.
For a discussion of positive operators, and many further references, see
W. WALTER, Differential- und Infegral-Ungleichungen, Springer, Berlin, 1964.
4. Thcsc tcchniqucs havc bccn dcvelopcd in somc joint work with S. A/cn.
162 QUASILINEARIZATION AND NONLINEAR BOU NDARY-VALUE PROBLEMS
5. For a detailed discussion of routing problems, see the book by Bellman
and Dreyfus referred to above, as well as
R. KALABA, "On Some Communication Network Problems," Combinatorial
Analysis, American Mathematical Society, Providence, R. I., 1960.
M. POLLACK and N. WIEBENSON, "Solutions of the Shortest-route Problem-A
Review," Operations Res., Vol. 8, 1960, pp. 224-230.
D. L. BENTLEY and K. L. COOKE, "Convergence of Successive Approximations
in the Shortest Route Problem," J. Math. Anal. Appl., to appear.
8. See
R. BELLMAN, "On the Approximation of Curves by Line Segments Using
Dynamic Programming," Comm. ACM, Vol. 4, 1961, p. 284.
R. BELLMAN, B. GLUss, and R. ROTH, "On the Identification of Systems and
the Unscrambling of Data: Some Problems Suggested by Neurophysiology,"
--, "Segmental Differential Approximation and the 'Black Box' Problem,"
R. BELLMAN, H. KAGIWADA, and R. KALABA, Dynamic Programming and an In-
verse Problem in Neutron Transport Theory, The RAND Corporation,
RM-4495-PR, March 1965.
Appendix One
MINIMUM TIME PROGRAM
C
C
C
C
TI< 4J EC TOR Y
FOR lULl" t:QUATION 1.+ IY' )**2 + Y Y"
COMMON r,';,C,K
DIMENSION r(500),CIIO)
20 READ lO,XA,YA,XB,YH,H,NHP,NFP,L
10
PRINT 100,XA,YA,X8iYB,H,NHP,NFP,L
O.
100 FORMATIIH15X5HINPUTI IH05X23HXA,YA,XB,YB,H,NHP,NFP,L I
1 IH05X 211P2E14.4,3X),El4.4, OP3I6)
C
C
C
C
C
C
"IH= I XA-XB) IH
NN=NH/NHP
L=HIGHEST APPROXIMATION
IXA,YA), INITIAL POINT
IXB,Y8), FI"IAL POINT
MAX=L+l
Cll)=IYB-YA)/{XB-XA)
KM=O
PRINT 103,KM,CIl)
103 FORMATIIH048X,I2,lOX 2HC=,lPi14.6)
C
C
DO 3 K=2,MAX
N=2*K+4
DO 30 1=1,500
30 TIll=O.O
T(2)=XA
TI 3) =t-j
KM=K-l
DLl 1 I=l,KM
J=2*I+2
TIJ)=YA
1 TIJ+ll=CII)
6 TIN-2)=0.
C
TI N-l) =0.
TIN) =0.
TlN+l)=l.
TIN+2)=1.
TIN+3)=0.
PRINT 104
104 FORMATIIH142XIHX,20X4HUIX))
C
C
9
8
106
C
CALL INT(T,Nol,O.,O.,O.,O.,O.,O.)
DO 8 I=l,NN
00 9 J=l,NHP
CALL INTM
PRINT l06,TIZ),TIN-4)
FORMATI29XIP2E21.6)
APPENDIX ONE
C
260<;15 TRAJECTORY
CIK)=IYB-TIN-2)-YA*TIN+2/TIN)
3 PRINT 103,KM,CIK)
C FINAL APPROXIMATION AND EXACT SOLUTION
XI=0.S*IXH+XA+IYB**2-YA**2)/IXB-XA)J
RS=YB**Z+IXB-XI)**2
PRINT 105, XI,RS
lOS FORMATIIHI28XIPlE21.6)
C
C
K=MAX+l
N=2*K+4
OU 31 1=1,500
31 TlI)=O.O
Tl 2)=XA
T(3)=H
DO 11 I=I,MAX
J=2*1+2
TlJJ=YA
II TIJ+I)=CIl)
C
TIN-Z)=O.
TlN-lJ=O.
T( NJ=O.
Tl N+I )=1.
TIN+2)=I.
TIN+3)=0.
PRINT 101
101 FORMATIIHI42XIHX,ZOX4HUIXJ,15X8HEXACT UJ
C
C
19
2
102
C
107
C
CALL
)(=XA
F=NFP
O=F*H
"lNN=NH/NHP
DO 2 I = I, NNN
00 19 J=I,NFP
CALL INTM
)(=X+O
Y=SQRTFIABSFIRS-I)(-XIJ**2JJ
PRINT 102,TI2J,TIN-4),Y
FURMATI29X,IP3E21.6J
CIKJ=IYB-TIN-2)-YA*TIN+Z) J/TIN)
CA={XI-)(A)/YA
PRINT 103,MAX,CIKI
PRINT 107.CA
FORMATIIH060X2HC=,IPEI4.6)
GO TO 20
ENOII,I,O,O,O,O,l,O,O,O,O,O,O,O,O)
165
C
260'>13
OAUX
SUilROUTI:,E DAUX
T,\j,C,K
DlHENSION TISOO),CI10)
TlN+4)=Cll)
f(,'H5):0.O
DO ) 1=2, K
J=2*I+2
"1:HJ
TCM)=TIJ+l)
4A=-2.*TIJ-l)/TIJ-2)
88=1 1.+TIJ-l )**2)/T(J-2)**2
CC=-2./TIJ-2)
TIM+l):AA*TIJ+I)+RB*TIJI+CC
T( M+2) =TI J+31
TIM+31=AA*TIJ+31+8B*TIJ+21
TIM+4)=TIJ+'j)
TIM+5)=AA*TIJ+51+RB*TIJ+41

ENDII,I,O,O,O,O,I,O,O,O,O,O,O,O,O)
Appendix Two
DESIGN AND CONTROL PROGRAM
168 QUASILINEARIZATION AND NONLINEAR BOUNDARY-VALUE PROBLEM
260550 DESIGN PROBLEM
COMMON NI,KMAX,NP,NMAX,HGRID,CX,T,X,y,U,V,W,PREV,H,P,A,8.C.HE
DIMENSION
PREVIIlI
C X. STATE VARIA8LE
C Y, CONTROL VARIA8LE
C UI=MU1, LAGRANGE MULTIPLIER
C VI=AI, DESIGN PARAMETER
C
C
I CALL INPUT
2 CALL START
C K ITERATIONS
C
3 00 19 K=I,KMAX
C INITIALIZE
00 5 1=1,250
5 TI I 1"'0.
Tl21:0.
Tt 31=HGRID

Tt91
z
1.
TIIIlI:I.
Tt 19J"'I.
NzO
PRINT III,
NEQ=20
CALL INTIT,NEQ,Nl,O.,O.,O.,D.,O.,O.1
C INTEGRATE P'S AND H'S, STORING GRID VALUES
C
C
N .. O
8 00 II M= I , NP
NaN+1
)(zWCI,NI
,(=WI2,NI
UzWC3,NJ
VzWCll,NI
CAll INTM
L=3
00 9 1=1,1l
00 9
lzL+I
9 HII,J,NI=TILI
00 10 JzI.1I
L=L+I
10 P(J.N)=TILI
II CONTINUE
6 PRINT
PRIHT .... .1.1..1--.t-l--t (.HI t. JLNI. J'" 1.4J ... jzl. 111 .. 1 PI J.N). J=I. III
III
IFIN-NMAXI8.7.7
7 PRINT IID.K
FORMATIIHO/6SX9HITERATION.131
C EVALUATE CONSTANTS
C
APPENDIX TWO
C
C
260550 DESIGN PR08LEM
NaNMAX
AII.1J"'HI2.2,NI
AI2.1,=HI2,3,NI
Al1,21"'HI3,2,NI + HI4,2,NI
A12,2J"'H13,3,NI + HI4,3,NI
BII'=-PI2,NI - CX*HII.2,NI
8121=-pel,NI - CX*HII,3.NI
OET-AI 1,11*A12,21 - AI2,II*AI 1,21
G=ABSFIOETI
IFIG-.0000011200,200, 12
C
200
201
PRINT 201,OET.IIAII,JI.J=I,2I,8IJJ,I=1.21
FORMATIIH020X12HOETERMJNANT=,EI6.6/r40X3E20.611
CALL EXIT
C
c
12 ClI'=CX
CI21-18111*AI2,21 - B12J*A11,211/0ET
Cl31=18121*AII,11 - 81 II*AI2,ll Ii0ET
C(41=Cl31
GRIO=O.
PRINT 45, GRIO,ICIJI,J=l,41
45 FORMATrlH026X4HGRIO,I4X1HX,19X1HY,ISX2HMU,19XIHAI
20XF10.4,4E20.61
C NEW VALUES OF VARlA8LES
C
N=O
13 DO 14 Mal,NP
N=N+1
DO 14 1= 1,4
WI hNI"'PI I,NI
00 14 J-I,4
+ CIJI*HIJ,I,NI
FN=N
GRJD=FN*HGRJO
PRINT 50, GRIO,IWII.NI,J=I.41
IFIN-NMAXI13,15.15
C COMPARE C'S, PRESENT AND PREVIOUS
C
c
15 DO 16 lal.4
G=A8SFICII'-PREVIIIJ
I 1116,16,11
16 CONTINUE
GO TO I
11 DO IS I = I. 4
18 PREvel'=ClII
19 CONTINUE
GO TO 1
50 FORMATI20XFI0.4,4E20.61
ENDII,I,O,O,O,O,I,O,O,O,O,O,O,O,OI
169
170 QUASILINEARIZATION AND NONLINEAR BOUNDARYVALUE PROBLEM
260552 INPUT - DESIGN
SUBROUTINE INPUT
COMMON Nl.KMAX,NP,NMAX,HGRIO,CX,T.X.Y,U,V,W,PREV,H,P,A,B,C
DIMENSION

C Nl. INTEGRATION UPTlOJIi
C KMAX, NUMBER OF ITERATIONS
C NP. NUMBERQf GRIDS BEFORE PRINT-QUT
C NMAX, TOTAL NUMBER OF INTEGRATIONS IN INTERVAL
C HGRID. GRID SIZE
C ex, X(O)
C
READ 110,Nl.KMAX,NP,NMAX,HGRIO,CX
PRINT
110
10 PRDBLMJI
1
2
RETURN
END(1.1.0."O.0.O.1.0.0.0.0.0.0.0.0)
APPENDIX TWO
17
6055) START (FOR 260550)
<;IJBRntITJ NF START
COMMON Nl,KMAX,NP,NMAX,HGRIO,CX,T,X,Y,U,V,W,PREV,H,P,A,B,C
DIMENSION

INITIALLY T=O. X(T)=CX. YIT1=0. MUIll=O. AII1=0
DO 1 N= 1 ,NMAX
WCl.Nl=CX
WI2,N)=0.
WU,N)=O.
W(IJ,N)=O.
00 2 1=1,
2 PR<.:VIIl=WII,l)
K=O
PRINT
GRIO=O.
PRINT
FORMATllHO/65X9HITERATION,I31

11H019X22HFOR ALL VALUES OF
RETURN
ENO(1.1.0.0.0.0.1,0.0.0.0,0,0.O.O)
60553 DAUX IFOR DESIGNI
SUBROUTINE DAUX
COMMON
DIMENSION
PREVt",
l=20
00 I 1=1,11
l=l+4
Ttll=-V-Ttl-201 + TIl-191 -X-Tll-111
Tll+II=T(l-201 + V-T(l-191 + Y_T(l-ITI
TIl+2J=Y-TIl-201 + X-TIl-IYI
Tll+31=O.
TI401=-V-T(201 + TI211 - X-T1231 + V-X
Tl4 1)=f(20 I + V*TI 211 + Y-Tl231 -- V-V
T(421=Y-TI201 + X-Tt211 - X-y
TI"31=O.
RETURN
END(I.l.0.0.0,O,l.0.0,O.O.O,O,O.OI
Appendix Three
PROGRAM FOR RADIATIVE TRANSFER-INVERSE
PROBLEM FOR TWO SLABS, THREE CONSTANTS
$JOB 26G9,STRAT3,HK016u,5,0,20,P
$PAUSE
IIBJOR STRAT2 MAP
IIBFT( RTINV
(
(
(
(
(
C
2
899
1000
90\.1
1001
901
(OMMON N,RT(7),WTI7),WRI7),ARI7,7),NPRNT,M1MAX,KMAX,DELTA,XTAU
1 XLAMI3h B217t7),RZI7,7).JFLAG,RI28dC1hTI14911,SIG,
Z PI28d01)tfiIZ8,3,101),PLAMI3),HLAMI3,3),P217,7),
3 H217,7,3),(0'lSTI3hNEQ
PHASE
READ1000,N
PRINT899
PRINT900,N
READ1LJ01, I RT (1),1=1 ,N)
PRINT901,IRTII) ,1=1,N)
READ1v01,IWT(1),1=1,N)
PRINT901,IWT(1),1=1,N)
DO 2 l=l,N
WR I I) =WT I I)/RT 11 )
DO 2 J = 1, N
ARII,J)= 1.0/RTII) + 1.0/RTIJ)
FORMATl1H146X36HRAD1ATIVE TRANSFER - INVERSE PROBLEM I )
FORMAT(6112)
FORMAT(6120)
16E1Z.8)
16"ZO.8)
READ1uOO,NPRNT,M1MAX,KMAX
PRINT900,NPRNT,M1MAX,KMAX
READ1J01,DELTA
PRINT901,DELTA
READ1J01,XTAU,IXLAM(1) ,1=1,3)
PRINT902
PRINT903,XTAU,IXLAMII) ,1=1,3)
902 FORMATl1H1Z3HPHASE I - TRUE SOLUTION II
903 FORMATI1HCI
1 1X11HTHICKNESS =, F10.4
Z 1XllHALBEDOIX) =, 20HA + B*TANHl10*IX-C II
3 1X3HA =, E16.8, 10X3HB =, E16.81 10X3HC =, E16.8 II)
CALL NONLI N
DO 3 1=1, N
DO 3 J=l,N
3 B211,J)=RZII,J)
C PHASE II
C
C
4 ,1=1,3)
K=O
PRINT904,K
PRINT903,XTAU,IXLAMII) ,1=1,3)
CALL NONLIN
APPENDIX THREE
(
(
(
(
(
(
(
904 FORMAT I 1H 1 13HAPPROXIMATION, 13/ )
QUASILINEARIZATION ITERATIONS
DO 5 K1=1,KMAX
PRINT904,K1
CALL PANDH
(ALL LINEAR
(ONTINUE
RI:AD1
0
00.IGO
GO TO 11,4).I GO
END
$IBFT( DAUX LIST
SUBROUTINE DAUX
DIMENSION V2(707),XI3).FI7),GI7)
1 .VLAM(3)
17
(OMMON N,RT(7),WTI7),WRI7),ARI7,7),NPRNT,M1MAX,KMAX,DELTA,XTA
1 XLAM(3), B217,71 ,R217,7) ,IFL/IG,RI28olC1) ,T(1491) .SIG,
2 PI28,101),HI28,3,101),PLAMI3),HLAMI3,3),P217,7),
3 H217,7,3),(ONSTI3),NEQ
GO TO 11,2),IFLAG
(
(NONLINEAR
C
(
(
L=3
DO 4 l=l,N
DO 4 J=l.I
L=L+1
4 V21I,J)=TIl)
DO 5 l=l,N
DO 5 J=I,N
5 V21IoJ)=V2IJoI)
DO 51 1=1.3
L=L+1
51 VLAMII)=TIL')
SIG=T(2)
Y=XTAU*SIG
DO 52 1=1,3
52 XI I )=VLAMI I)
CALL ALBEDQIY,X,Z)
ZLAMDA=Z
DO 6 l=l,N
FII)=O.O
DO 7 K=l,N
7 FII)=FII) + WRIK)*V21I,K)
6 FII)=G.5*FII) + 1.0
1)0 8 l=l,N
DO 8 J=l,1
176 QUASI LINEARIZATION AND NONLINEAR BOU NDARYVALUE PROBLEMS
(
L=L+l
DR=-AR(I.J)*V2(I.J) + ZLAMDA*F(I)*F(J)
8 T(l)=DR
DO 9 1=1.3
L=L+l
9 T(l)=O.O
RETURN
(
(LINEAR
(
(
(
2 SIG=T(2)
Y=XTAU*SIG
DO 21 1=1.3
21 X(I)=XLAM(I)
(ALL ALBEDO(Y.X.Z)

DO 16 1=1. N
F ( I ) =0. ()
DO 17 K=loN
17 F(I)=F(I) + WR(K)*R2(I.K)
16 F(I)=O.5*F(I) + 1.0
(P I S
(
(
L=3
DO 14 l=l.N
DO 14 J=l.I
L=L+l
14 V2(I.J)=T(L)
DO 15 1=1. N
DO 15 J=I.N
15 V2(I.J)=V2(JoI)
DO 18 1 = 1 3
L=L+l
18 VLAM ( I ) = T ( l)
DO 10 1=1. N
G(I)=O.O
DO 10 K= 1. N
10 G(I)=G(I) + (V2(I.K)-R2(I.K))*'tJR(K)
ARG=10.C*(Y-XLAM(3) )
TARG=TANH(ARG)
XTANX=-10.0*XLAM(2)*(1.0-TARG**2)
3 + NEI"J
DO 12 1=1. N
DO 12 J=1.I
FIJ=F(I)*F(J)
(APF=-AR(I.J)*R2(I.J) + ZLAMDA*FIJ
Tl=(APF

1 + + F(J}*G(I}}
T3= ( 1) -XLAM (1) } *F I J
APPENDIX THREE
(
T4=IVLAM(2)-XLAMI2) )*TARG*FIJ
T5=IVLAM(3)-XLAMI3))*XTANX*FIJ
M=M+1
12 TIM)=T1+T2+T3+T4+T5
DO 19 1=1,3
M=M+1
19 TIM)=O.O
(H'S
(
(
(
(
(
DO 100
D0241=1.N
DO 24 J=l,1
L=L+1
24 V211.J)=TIl)
DO 25 l=l.N
DO 25 J=I.N
25 V211.J)=V2IJoI)
DO 26 1=1.3
L=L+l
26 VLAMII)=TIl)
DO 20 l=l.N
GIl) =0.0
DO 20 J=l.N
20 GII)=GII) + V2(I,J)*WRIJ)
DO 22 l=l.N
DO 22 J=lt!
FIJ=FII)*FIJ)
T1=0.0
T2=-ARII.J)*V211.J) + 0.5*ZLAMDA*IFII)*GIJ) + FIJ)*G(I))
T3=VLAM(1)*FIJ
T4=VLAM(2)*TARG*FIJ

1
22 TIM)=Tl+T2+T3+T4+T5
DO 29 1=1.3

29 TIM)=O.O
100 CONTINUE
RETURN
END
$IBFTC NONLI N
SUBROUTINE NONLIN
177
COMMON N.RT(7),WTI7).WRI7),ARI7.7).NPRNT.M1MAX,KMAX.DELTA.XTAU
1 XLAM(3), 6217,7) ,R217,7),IFLAG.RI28.1Cl).TI1491).SIG.
2 P 1 28 1 0 1 ) , H 1 28.3, 101 ) , P LA 1 3 ) H LA" 1 3 3) P 2 1 7 7)
, H217.7.3l.CONSTI,ltNEQ
C NONLINEAR D.E. FOR TRUE SOLUTION OR FOR INITIAL APPROX.
(
IFLAG=l
C
C
C
C
T(2)=0.0
T(3)=DELTA
M=1
L1=0
L'l='l
DO 1 1=1,N
DO 1 J=1.I
L1=Ll+1
L3=L3+1
R211,J)=0.0
RILl,M)=R211,J)
T 1 L3) =R2 1 I ,J)
DO 2 1=1,3
L3=L3+1
2 TIL3)=XLAMII)
NEQ=IN*IN+l))/2 + 3
CALL INTSIT,NEQ,2,0,0,0,0,C,O)
SrG=T(2)
CALL OUTPUT
DO 5 M1=1,M1MAX
DO 4 M2= j, NPRNT
CALL INTM
M=M+l
L1=0
L3=3
DO 3 1=1,N
DO 3 J=l,1
L1=L1+1
L3=L3+1
R21 I ,J)=TIL3)
3 RI Ll,M)=R21 I,J)
4 SIG=T(2)
5 CALL OUTPUT
RETURN
END
$IBFTC PANDH
SUBROUTINE PANDH
COMMON N,RT(7),WTI7),WRI7),ARI7,7),NPRNT,M1MAX,KMAX,DELTA,XTAU
1 XLAM(3), B2(707),R217,7),IFLAG,RI28,JOl),TI1491),SIG.
2 PI28,lOl),HI28,3,101),PLAMI3),HLAMI3,3),P217.7),
3 H217,7,3)oCONSTI3),NEQ
IFLAG=2
T(2)=0.0
T(3)=DELTA
M=l
Ll=O
L3=3
DO 1 1=1, N
APPENDIX THREE 179
DO 1
Ll=Ll+1
L3=L3+1
P(Ll.MI=O.O
T(L3I=P(Ll.MI
DO 2

PLAM(II=O.O
2 T(L31=PLAM(II
c
CH'S
C
(
(
C
(
DO 7

DO 3
DO 3
Ll=Ll+l
L3=L3+1
H(Ll.K.MI=O.O
3 T(L31=H(Ll.K.MI
DO 7 1=10'3

HLAM( I.KI =0.0
IF(I-KI7.6.7
6 HLAM(I.KI=l.O
7 T(L31=HLAM(I.KI

DO 8 l=l.N
DO 8 J=l.r

8 RZ(I.JI=R(L.MI
DO 9 l=l.N
DO 9 J=I.N
9 RZ(I.JI=R2(J.II
NEQ=4*((N*(N+111/2 + 31
CALL INTS(T.NEQ.2.0.0.0.0.v.01
LMAX=(N*(N+111/2
P R I NT 52.r ( 2 I ( P ( L MI. H ( L 01 MI. L = 1 L MAX I
52 FORMAT(lHOF9.4.5E20.8/(10X5E20.81 I
DO 51
DO 50 M2=1.NPRNT
(ALL INTM
M=M+l
(pREV.APPROX. R(I.JI
Ll=O
DO 10 I=I.N
DO 10 J=l.I
Ll=Ll+1
10 R2(I.JI=R(Ll.MI
DO 11 l=l.N
DO 11 J= I , N
11 R2II,J)=R2IJ,I)
L1 =0
L3=3
DO 12 I=I,N
DO 12 J=I'[
L1=L1+1
L3=L3+1
12 PILl,M)=TIL3)
L3=L3+3
DO 13 K=I03
L1=0
DO 14 I=I,N
DO 14 J=I'[
L1=L1+1
L3=L3+1
14 HILl,K,M)=TIL3)
13 L3=L3+3
50 CONTINUE
51 PRINT52,T(2),IPIL,M),HIL,I,M),L=1,LMAX)
RETURN
END
BFTC LINEAR
SUBROUTINE LINEAR
DIMENSION CHKI(3)
DIM ENS ION A I 49,3) , B I 49 ) , E MA T I 50,50) , PI VO T I 50 ) , IN D E X I 50 ,
1,IPIVOTI50),FVECI50ol)
COMMON N,RT(7),WTI7),WRI7),ARI7,7),NPRNT,M1MAX,KMAX,DELTA,XTAU,
1 XLAMI3lo B217,7"R217,7),IFLAG,PI28olCl)oTI1491),SIG,
2 P128ol01) ,HI28,3ol01) ,PLAM(3) ,HLAMI3,3) ,P2(707),
3 H217,7,3),CONSTI3),NEQ .
OUNOARY CONDITIONS
MLAST=NPRNT*MIMAX + 1
DO 1 K = 1,3
L=O
DO 2 1=I,N
DO 2 J=I'[
L=L+l
H211,J,K)=HIL,K,MLAST)
DO 1 1=1,N
DO 1 J= I, N
H2II,J,K)=H2IJ,I,K)
L=O
DO 3 1=1,N
DO 3 J=1d
L=L+l
3 P2II,J)=PIL,MLAST)
DO 4 1=1,N
DO 4 J= I, N
4 P211,J)=P2IJ,I)
EAST SQUARES
DO 5 K=I,3
L=O
DO 5 1=1, N
APPENDIX THREE
(
(
(
(
(
C
(
DO 5 J=I.N
L=L+l
5 AIL,K)=H211,J.K)
L=O
DO 6 1=I,N
D06J=],N
L=L+l
6 BIl)=B21 I ,J) - P21 I,J)
LMAX=N**2
PRINT60
60 FORMAT(lHO)
DO 61 L=l,LMAX
61 PRINT82.IAIL.K),K=I,3).BIL)
DO 8 '1=1,3
DO 7 J=I03
SUM=O.O
DO 9 L=I,LMAX
9 SUM=SUM + AIL,I)*AIL,J)
7 FMATII.J)=5UM
SU"'=O.O

10 SUM=SUM + AIL,1)*BIL)
8 FVE(II,1)=5UM
PRINT60
D081 1=1.3
81 PRINT82,IEMATII,J),J=I03),FVE(llol)
82 FORMATII0X6E2v.8)
CALL MATINVIEMAT,3,FVEC,I,DETERM,PIVOT,INDEX,IPIVOT)
DO 11 1=1,3
11 CONSTII)=FVECllol)
DO 20 1=1,3
20 XLAMII)=CONSTII)
PR I NT 90 3. X T AU, I XLAM I I ) 1=1 ,3)
903 FORMATll>-iOI
1 lX11HTHICKNESS =, E16.8 I
2 lX11HALBEDOIX) =, 20HA + S*TANHII0*IX-()) II
lX3HA =, EI6.8, 10X3HB =. EI6.8, 10X3HC =, E16.8 II)
(NEW APPROXIMATION
(
M=1
L=O
D0121=I,N
DO 12 J=lo1
L=L+l
SUM=PIL.M)
DO 13 K=I03
13 SUM =SUM + CONSTIK)*HIL,K,M)
181
(
(
12 R(L."1)=SU"1
L=O
DO 14 I=l,N
DO 14 J=l.r
L=L+l
14 R2(I,J)=R(L.M)
SIG=O.O
(ALL OUTPUT
DO 50 M1=1,M1MAX
DO 18 M2=1,NPRNT
M=M+1
L=O
DO 15 I=l,N
DO 15 J=l.I
L=L+1
SUM=P(L."1)
DO 16 K=l03
16 SUM=SUM + (ONST(K)*H(L.K,M)
15 R(L.M)=SUM
L=O
DO 17 I=ltN
DO 17 J= 1, I
L=L+1
17 R2(I.J)=R(L,M)
18 SIG=SIG + DELTA
50 (ALL OUTPUT
RETURN
END
$IBFT( OUTPUT
SUBROUTINE OUTPUT
DIMENSION X(3)
(OMMON N.RT(7).WT(7),WR(7),AR(7,7),NPRNT.M1MAX,KMAX,DELTA.XTAU
1 XLAM(3), B2(7t7),R2<7t7).IFLAG.R(28tlC'1),T(1491)oSIG.
2 P(28tl01) ,H(28.3.101) ,PLAM(31 ,HLAM(3.3) ,P2(707).
3 H2(7.7.3),(ONST(3).NEQ
DO 1 1=1. N
DO 1 J= I. N
R2(I.J)=R2(J.r)
Y=XTAU*SIG
DO 3 1=1,3
3 X( I )=XLAM( I)
(ALL ALBEDO(Y,X,Z)
PRINT100. SIG.y,Z
100 FORMAT(lHO 7HSIGMA =,F6.2, 4X5HTAU =, F6.2, 4X8HALBEDO =.F6.21
DO 2 J=l.N
2 PRINT101.J. (R2( I ,J) .I=l.N)
101 FORMAT(I10. 7F10.6)
RETURN
END
$IBFT( ALBEDO
SUBROUTINE ALBEDO(Y,X.Z)
DIMENSION X(3)
APPENDIX THREE
(aMMON N,RT(7),WT(7),WR(7),AR(7,7),NPRNT,MIMAX,KMAX,DELTA,XTA
1 XLAM(3), B2(7,71,R2(707),IFLAG,R(28dOl),T(1491),SIG,
2 P(28,101),H(28.3,101),PLAM(3),HLAM(3,3),P2C707),
3 H 2 ( 7 07 , 3 ) , (0 N 5 T ( 3 "N EQ
ARG=1 0 .O*(Y-X(3))
Z=X(l) + X(2)*TANH(ARG)
RETURN
fND
$ENTRy RTINV
7
25446046E-0112923441E-0029707742E-0050000000E 0070292258E 008707655
97455396E 00
64742484E-Cl13985269E-0019091502E-0020897958E-0019091502E-001398526
64742484E-Jl
llJ 10 4
0.01
1.0
1.0
0.5
r.6
0.1
.09
Appendix Four
VAN DER POL EQUATION PROGRAM
(VDP021 VAN DER POL - PrlASES II, III
(OMMON
PREV,NOBS,IFLAG
DIMENSION
H(3,3,1001) ,P(3,1001) ,PREV(3)
(
( INPUT NOBS OdSERVATIONS OF X AT KNOWN
( DETERMINE SYSTEM PARAMETER LAMBDA (Z IN FURTRAN)
( INPUT AND START
(
(
(
(
(
(ALL INSTRT
K ITERATIONS
DO 99 K=l,KMAX
DO 2 1=1,150
T!I)=C.O
T(2)=TI
T(3)=rlGRID
T(4)=1.0
T(8)=1.0
T(12)=1.0
X=PREV(l)
U=pREV(2)
Z=PREV(3)
(ALL INT(T,12,N1,O.,0.,O.,O.,0.,0.)
N=l
L=3
DO 21 1=1,
DO 21 J=1,3
L=L+1
21 H(!,J,N)=T(L)
DO 22 1=1,3
L=L+1
22 P(!,N)=T(l)
( INTEGRATE P'S AND H'S
(
DO 4 N=2,NMAX
X=W(l,N)
U=\-I (2,N)
Z=\-I (3,N)
(ALL INTr-'.
L=3
DO 3 1=103
DO 3 J=1.3
L=L+1
3 H(r.J.N)=T(L)
DO 4 1=103
L=l+l
4 P(!.N)=T(l)
( DETERMINE (ONSTANTS. OR INITIAL VALUES
(ALL (NSTNT
TIME=TI
PRINT 50. K.TIME,(B(I,l) .1=1.3)
(
( NEW VARIABLES
DO 7 N=2.NMAX
DO 6 1=1.3
APPENDIX FOUR
187
6
7
C
C
8
9
1 ",
C
99
C
50
70
"II.NI=PII.NI
D06J=1.3
VI 1 I N 1 ='tI 1 I. N 1 +8 1 J.1 1 *H 1 J. I N 1
FN=N-1

PRINT 70. TlME,(1J1 I .N) d=l >31
COMPARE CONSTANTS
DO 8 1=1.3
G=ABSF 1 B 11.11 -PREVI I II
IFIG-.COOJ01)8.8.9
CONTI NUF
GO TO 1
DO 1e 1=1 t3
PREY 1 I 1 =R 1 I tl 1
CONTINUE
GO TO 1

1
FORMATI30XF1G.2.3E20.61
END
CVDP022 INPUT-START. VAN DEi, POL PHASE Iltlll
INSTRT
COMMON
PREV.NOBS.IFLAG
D I ME NS I ON SCRACH 1 64 1 Til 5':; 1 N Ell CO 1 , X OBS 1 1001 , 'Ii 13.1001 1 ,
HI 3,3,10011 , PI 3 1 vO 1 1 A 1 16, 161 , B 1 16.11 P R EV (3)
C INPUT
C
C
C
C
READ ,XOBSIII,I=l.NOBSI
IFINMAXI9,9,1
PRINT .XOBSIII,I=l.NOBSI
START
IFLAG=l
READ
DO Z 1=1 t1 SO
TIII=O.O
TIZI=TI
TI31=HGRID
T 141 =X,
T151=U
T161=Z
CALL INT(T,3tNltO.,O.,o.,J.,C.,o.)
K=O
PRINT 50, K,TIZI,ITIII,I=4.61
DO 4
(ALL INH'
DO 31=1.3
3 WII,NI=TII+31
4 PRINT 70. T!2ltITIII.r=4.61
IFLAG=2
PREVI I I=X
PREVI21=L!
PRFVI31'Z
C
RETURN
9 CALl:. EXIT
10 FORMAT(lH130X
1 58HVAN DER POL - PHASE II - DETERMINATION OF SYSTEM PARAMETERII
2 20X4I10,FIJ.4/118XI12,E16.6,I12,Ell).6,I12,E16.6,I12,El6.61 I
50 FORMATIIHOI159X9HITERATION,I31138XIHT,13XIHX,19XIHU,17X6HSYSTEMII
1 30XFI0.2,3E20.61
70 FORMATI30XFI0.2,3E20.61
11 0 FORMA T 141 5 , F 1 V 2 , 2 ( 14, Ell. 6 II I 4 I 14, E 11 I) I I I
120 FORMAT(4E12.1)1
END
CVDP023 DAUX - VAN DER POL - PHASf 11,111
C
SUBROUTINE DAUX
COMMON SCRACH,T,Nl,NMAX,KMAX,HGRID,NTIME,x085,W,H,P,A,B,X,U,l,TI
PREV,NOBS,IFLAG
D I NS ION SC RA CH (64 I , TIl 501 .. NT! ME I lOa I , X OBS I 1 OJ I , W I 3,1001 I ,
HI3,3,10011,PI3,lJ011,AI16,161,8116,11,PREVI31
GO TO (102 I olFLAG
C NONLINEAR EQUATIONS
TI71=T(51
TI8'=-ITI41**2-1.I*TI51*TI61-TI41
TI91=(,.
RETUR"J
C LINEAR EUUATIONS
2 XX=X**2
AA=-2.*X*U*Z-1.
BB=Z*ll.-XXI
CC=U*ll.-XXI
L=13
DO 3 1=1,4
L=L+3
TILl=TIL-lll
TIL+ll=AA*TIL-121+BB*TIL-111+CC*TIL-101
3 TIL+21=O.0
TIL+ll=TIL+11+U*Z*13.*XX-1.1
RETURN
END
CVDP024 CNSTNT - PHASE II
C
SUBROUTINE CNSTNT
COMMON
PREV,NOBS,IFLAG
DIMENSION SCRACHI641,T 1150 I ,NTIMEI 100 I ,XOBSI 1001 ,WI 301001 I,
HI 3 , 3 , 1001 I , PI 3 , 1 00 1 I , A I 16,16 I , BIll), 1 I , PR EV 131
DO 1 1=1,3
N=NT!MEIII
Blloll=XOI3SIII-Pll,NI
DO 1 J=1,3
AI [oJI=HIJo1,NI
CALL MATINVIA,3,B,1,DETI
DO 2 1=1,3
2 Wllo1l=Bllo1l
RETURN
END
Appendix Five
ORBIT DETERMINATION PROGRAM
C260582 CELESTIAL MECHANICS - II
C

NGRID,THETA,W,H,P,A,A,X,U,Y,V
DIMENSION
.P(4,580).A(16.16).B(16.1)
1 CALL INPUT
2 CALL START(W)
C K ITERATIONS
DO 19 K=l.KMAX
NEQ=20
4 T(2)=TZERO
T(3)=HGRID
DO 5 1=4.43
5 T(I)=O.
T(4)=1.
T(9)=1.
T(14)=1.
T(19)=1.
6 CALL INT(T.NEQ.N1.A2.A3.A4.A5.A6.A7)
C INTEGRATE OVER RANGE
7 DO 11 N=l.NMAX
X=W( loN)
U=W(2.N)
Y=W(3,N)
V=W(4.N)
CALL INTt.!
C STORE P'S AND H'S
C
C
8
9
10
11
12
13
114
14
15
115
C
C
40
16
L=3
DO 9 1=1.4
DO 9 J=1.4
L=L+1
H( I.J.N)=T(l)
DO 10 J=1.4
L=L+1
P(J.N)=T(L)
CONTINUE
CONSTANTS
DO 14 1=1.4
N=NGRID( I)
THET=THETA(i)
S THET=S I NF (THET)
CTHET=COSF(THET)
"0 13 J=1.4
A(I.J)=H(J.1.N)*STHET -H(J.3.N)*CTHET
B( 1.1 )=( 1.-P( loN) )*STHET +P(3.N)*CTHET
FORMAT(lH09X5E20.6)
PRINT 114. (A( I,JJ) ,JJ=1.4) .B( loll
CALL MATINV(A.4.B.l.DET)
PRINTl15,(B(I.1).I=l,4)
FORMAT(lH09X4E20.6)
COMPUTE NEW 'II'S
PRINT40.K
FORMAT(lHO/65X 9HITERATION.13//26X4HGRID.14X1HX.19X2HX'.18XIHY
19X2Hy' )
DO 18 N=l.NMAX
DO 17 1=1.4
W( I.N).P( I.N)
APPENDIX FIVE 191
DO 17 J=1.4
17 + [](J.ll*H(J.J.N)
50
18 PRINT50.N,(W( I.N) 01=104)
(
19 CONTINUE
C
GO TO
END
(26C585 INPUT
C
SUBROUTINE INPUT

NGRID.THETA,W.H.P,A.B.X.U.y,V
DIMENSION
, P (4.5"0) A (16.1"6) , g (16.1 1

110 FORMAT(313.7E8.1)
REA D1 20 ( N G RID ( I 1 , T H ET A ( I ) I = 1 .4 1
120 FORMAT(4(13.EI2.4)I
PRINT10,Nl,NMAX,KMAX.HGRID,A2.A3.A4,A5.A6.A7,(NGRID(I).THETA(I
11=104)
10 FORMAT(IHI2UX 8HINPUT I
1
? 12X2HA5.12X2HA6,12X2HA7/15X315.7EI4.41/
'21X4HN(I).10X8HTHETA(I).6X4HN(21.10X8I-'THETA(2).
4 6X4HN(31.1
u
X8HTHETA(31.6X4HN(41.10X8HTHETA(4)/15X4(110.EI8.6

END
(260584 DAUX
(
SU8ROUTINE DAUX
(OMMON S(RA(H.T.NEQ.Nl.NMAX,KMAX,HGRID.A2.A3.A4.A5.A6.A7.
NGRID.THETA.W,H,P,A,B.X,U.y,V
DIMENSION S(RACH(64).r (250) ,NGRID(41 .THETA (4) .'J( 4,500) .H( 4.4.
,P(4,500l.A(16,161,B(16.11
XX=X**2
YY=Y**2
'<=XX+YY
S=R**3
D=SQRTF(R**2*S)
E=SQRTF(S)
AA=3.*X*Y/D
BB=(2.*X)l-YY)/D
((=(2.*YY-XX)/D
N=23
,,0 1 1=1.4
N=III+1
T(N)=T(N-191
T(N+l)=T(N-20)*B8 + T(N-18)*AA
T(N+2)=T(N-17)
T(N+3)=T(N-20)*AA + T(N-181*((
N=N+3
T(40)=T(211
T(41)=T(20)*88 + T(221*AA - 3.*X/E
T(421=T(23)
T(43)=T(20)*AA + T(22)*(( - 3.*Y/E
RETURN
END
(260583 START(W)
SU9ROUTINF START(W)
COMMON
NGRID.THFTA.W.H.P.A.B.X.U.y.V
D 1 NS 1 ON SCRACH ( 64 ) T ( 25 C ) NGR 1 D ( 4) THf' T A (4) W ( 4.500) H ( 4.4
P(4.50
n
).A(16.16).B(16.1)
C X=l. X=O. y=O. Y'=C
c
K=O
PRINT4r:.'(
40 65X
1
DO 2
\<i(l.r-n=l.
no 1 1=2.4
1 ';1( T ,N) =f"I.
2 PRINT50."."I(I.N)d=1.4)
5C FORMAT(2CX.I10.4E20.6)
RETURN
E"D
Appendix Six
CARDIOLOGY PROGRAM
194
QUASILINEARIZATION AND NONLINEAR BOUNDARY-VALUE PROBLEMS
(MAIN2
(
HEART MODEL - INVERSE PROBLEM - 2
(
(
(
(
(
(
(
(
(
(
(
(
(
(
(
(
(OMMON NMAX,DELTA,TAU(5),SIGMA(5)
1 XOB S VR ( 4,3) , AD 1ST N ( 3 ,5) , F ( 5 , (31) , BOBS ( 3 ,81 ) .T I ME ( 81 ) , T ( 1323 )
(OMMON KMAX, ASIGMA(5loAKAPPA!5),P(5,81),H(5ol0,81),PS(5lo
1 HS (5,10) .FPREV( 5) ,AMATRX (243010) ,BVE(TR (243) ,EMATRX( 50,50).
2 FVE( TR ( 50) , (( 1 0 ) , V ( 3 ,81 ) , N(HANG, I (HA NG ( 5 ) , F (H,\NG ( 5 ) , I FL AG
,SCHAt\lG(5)
DISTANCE MODEL
5 DIPOLES OR CHAMBERS
GAUSSIAN FUN(TION FOR MAGNITUDE OF MOMENTS
TAU,SIGMA,ZKAPPA = TIME OF HALF-WIDTH, MAX. MOMEN
D(OS(I,J) = DIRE(. (OSINE OF DIPOLE I, AXIS J
XCHAMB(I,J) = LOCATION OF DIPOLE I, (OORDINATE J
XOBSVR(I.J) = LOCATION OF I, (OORDINATE J
J=1 X-AXIS OR X-DIRECTION
J=2 Y-AXIS OR Y-DIRECTION
J=3 Z-AXIS OR Z-DIRE(TION
(ALL FPST
READ1IJOO.NMAX
PRINT899
PRINT90Q,NMAX
READI001.DELTA
PRINT901.DELTA
( INPUT DIPOLE PARAMETERS
(
(
C
DO 2 N=1,5
2 READ1 002,I,TAU(I),SIGMA(I),ZKAPPA(I)
DO 3 N=1,5
3 READ1 v 03ol'(D(OS(I,J)oJ=1,3).(X(HM1B(I,j) .J=1.3)
PRINT902
DO 4 1=1,5
4 PRINT903,I,TAU(I),SIGMA(I),ZKAPPA(I),(X(HAMB(I,J).J=I,3),
1 (D(OS(I.J),J=l03)
( INPUT OBSERVER LOCATIONS
( (OMPUTE DIPOLE-OBSERVER (OEFFICIENTS
(
DO 5 N=I,4
5 READ10030l ,(XOBSVR( I,J )oJ=1,3)
DO 6 1=1,5
(ONUM=O.O
CODEN=O.O
DO 61 L=I03
YX=XOBSVR(4.L) - X(HAMB(I.L)
(ONUM=(ONUM + D(OS(I,L)*YX
61 (ODEN=(ODEN + yX**2
APPENDIX SIX
C
C
C
CODEN=SQRTCCODEN**3)
CZEROI=CONUM/CODEN
DO 6 J=1.3
CJNUM=O.O
CJDEN=O.O
DO 62 L=1.3
YX=XOBSVRCJ.L) - XCHAMBCI.L)
CJNUM=CJNUM + DCOSCI.L)*YX
62 CJDEN=CJDEN + YX**2

6 ADISTNCJ.I)= CJNUM/CJDEN -
PRINT904
DO 7 J=1.3
7 P R I NT 905 J C X OB S V R C J L ) L = 1 .3 ) C A,) 1ST N C J I ) I = 1 .5 )
J=4
PRINT905.J.CXOBSVRCJ.L).L=1.3)
C INPUT CHANG"'S CSIGM.'. KftPPA)
C
C
70 PRINT908
READIOOQ.NCrlANG
PRINT900.NCHANG"
IFCNCHANG.LE.O)GO TO 80
READIOOO. C ICHANGC 1).1=1 .NCHMG)
PR I NT 9 00. C I CHANG C I ) 1=1 NCHANG )
READ1 G01.CSCHANGCI).FCHANGCI).1=1.NCHANG)
P R I NT 9 01. C SCrlAN G C I ) FCHA N G C I ) 1=1 N CHA NG )
C STORE CHANGES
C
C
C
C
DO 71 l=l.NCrlANG
L=ICHANGCI)
SSIG=SIGMACl)
SIGMACL)=SCHANGCI)
SCHft.NG C I) =SSI G
SKAP=ZKAPPACL)
ZKAPPA C l) =FCHANG C I )
71 FCHANGCI)=SKAP
C INITIAL INTEGRATION STEP
C
8e IFLAG=l
K=l
L=3
DO 8 1=1.5
L=L+l
8 TCL)=ZKAPPACI)*EXPC-O.5*CTAUCI)/SIG,"'IICI))**2)
DO 81 1=1.5
L=L+1
81 TCl)=SIGMAC I)
195
T(2):0.0
T(3):DELTA
(ALL INTS(T.10.2.0.0.C.0.0.0.0.O,0.o.c.r)
L:3
DO 82 1:1.5
L:L+1
82 F( loK):T(l)
( OBSERVATIDN
(
(
(
(
(
DO 83 j:103
BOBS(j.K):O.O
DO 831:1.5
83 BOBS(j.K):BOBS(j.K) + ADISTN(j.I)*F(I.K)
TIME(K):T(2)
PRINT906
PR I NT 907. T I ME ( K ) ( F ( I K) I: 1 .5) ( BOBS ( j K ) j: 1 .3)
INTEGRATE
DO IOU K:2.NMAX
(ALL I NTM
L:3
DO 91 I: 1.5
L:L+1
91 F(I.K):T(L)
DO 92 j:103
BOBS(j.K):O.O
DO 92 I: 1.5
92 BOBS(j.K):BOBS(j.K) + ADISTN(j.I)*F(I.K)
TIME(K):T(2)
P R I NT 907. T I ME ( K ) ( F ( I K) I: 1 .5) ( BOBS ( j. K ) j: 1 .3)
100 (ONTINUE
(
(
( PHASE II - DETERMINATION OF HEART PARAMETERS
(
(
(
(
(
(ALL PHASE2
IF(N(HANG.LE.O) GO TO 102
( RESTORE ZKAPPA
DO 101 1:1.N(HANG
L:I(HANG(I)
SIGMA(L):S(HANG(I)
101 ZKAPPA(L):F(HANG(I)
(
( NEXT PROBLEM
102 READ100C.IGO
PRINT900,[GO
GO TO(1.70),[GO
(
APPENDIX SIX
C
C FORMATS
C
1000 FORMATI611Z}
lOCI FORMATI6E1Z.6}
899 FORMATIIH1 4X38HHEART MODEL - -PHASE I II}
FORMATI61Z0}
901 FORMATI6EZO.6}
1002 FORMATII10.3FIO.4}
1003 FORMATII1C,6F1C.4}
197
902
1 3X7HY-COORD,3X7HZ-COORD.5X.4X 6HX-DCOS.4X6HY-DCOS,
2 4X6HZ-DCOS II}
903 FORMATI15.SX3FIO.4,SX3F10.4,5X3FIO.4}
904 .5X. 3X7HX-COORD.3X7HY-COORD.3X7HZ-COORD.5
1
905 FORMATII10,5X3FIO.4. SX5FIZ.6}
906 FORMATIIH08XlrlT,5X. 5X5HFIIT}.5X5HFZIT}.5X5HF3IT} ,5X5HF4IT}.
I 5X 5 HF 5 IT} .5 X t7 X 5HV 1 IT} t7X 5 HV 2 IT} ,7 X 5 HV 3 IT} I I }
907
9U8 FORMATIIHI 4X13HABNORMALITIES II}
END
$IBFTC PHA5EZ
C
SUBROUTINE PHASE2
COMMON
1 XOBSVRI4t3} .ADISTNI3.5} .FIS.81} .BOB513.81} .TlMEI81}.TI13Z3}
COMMON ASIGlv,AI5}.AKAPPAIS}.PI5,81},HI5tlO.8I}.PSI5}.
I HS I 5 .10 } FPREV IS} AM A T RX I Z 4 3.10 } BVE CT R I 243 } EM A TR X I 50. 50}
2 FVECTRI50}.CIIC}.VI3.81}.NCHANG,ICHANGI5},FCHANGI5}.IFLAG
C SOLVE THE INVERSE PROBLEM
C
C ASIGMA APPROX. TO SIGMA
C AKAPPA APPROX. TO ZKAPPA
C PII,K} PARTICULAR SOLU. DIPOLE I, TIME K IMOMENT}
C HII,J,K}=HOMOGENEOUS SOLU. J, ...
C PSII} PARTICULAR SOLU. DIPOLE I ISIGMA COMPONENT}
C HSII,J}= HOMOGENEOUS SOLU. J, ...
C VIJ,K} = POTENTIAL DIFFERENCE OBSERVED BY ELECTRODE J
C AT TIME K, AS COMPUTED IN PHASE II. COMPARE
C WITH BOBSIJ.K} OF PHASE I.
C
C
C
PRINT899
READI000,KMAX
PRINT900,KMAX
PRINT904
DO 1 N=105
READIOOZ, I ,ASIGMAII} ,AKAPPAII}
PR I NT 905, I , AS I GMA I I } , AKA P PA I I }
C INITIAL APPROXIMATION
KI=O
PRINT909,Kl
(ALL APPRXI
(
(
HIGHER APPROXIMATIONS
(
IFLAG=2
DO 200 K1=I,KMAX
K=1
DO 2 1=1,5
P(!ol)=O.O
PS ( I ) =0.0
DO 3 J=lol0
H( I ,Jol) =0.0
3 HS(I,J)=O.O
H(!.I,1l=1.0
15=1+5
2 HS ( I , I 5 ) = 1 .0
L=3
DO 4 1=1,5
L=L+l
4 T(L)=P(!ol)
DO 5 1=10 5
L=L+l
T(L)=PS(I)
DO 6 J= 1010
DO 7 1=1,5
L=L+l
7
T ( L ) =H ( I , J , 1 )
DO 6 1=1,5
L=L+l
6 TlLl=HS(!,J)
T(2)=0.0
T(3)=DELTA
DO 8 1=1,5
8 FPREV(!)=F(i,K)
(
(ALL INTS(T,110,2,0.O,0.0,0.0,0.0,0.0,0.0)
(
DO 9 K=2,NMAX
CALL I NTM
L=3
DO 10 1=1,5
L=L+1
10 P(!,K)=T(l)
L=L+5
DO 11 J=I,10
DO 12 1=1,5
L=L+l
12 H(!,J,K)=T(L)
11 L=L+5
DO 9 1=1,5
9 FPREV( I )=F( I ,K)
C COMPUTE NEW APPROXIMATION
(
PRINT909,Kl
APPENDIX SIX
200 CALL APPRX2
C
C
RETURN
899 FORMAT(lHl 4X 8HPHASE II II)
1000 FORMAT(6112)
900 FORMAT(6120)
1002 FORMAT(IIC.10X.5F10.4)
905 FORMAT(IIO.2F10.4)
904 FORMAT(lHC.8X1HI.5X5HSIGMA.5X5HKAPPAII)
909 FORMAT(lH156X13HAPPROXIMATION.1311)
END
$ I BFTC APPRX 1
SUBROUTINE APPRX1
199
COMMON N.MA X DEL T A. T AU ( 5) S I GM A ( 5 ) ZKAP PA ( 5 ) DCOS ( 5.3) X CH AMB ( 5 .3
1 XOBS VR ( 4 3 ) A DIS T N ( 3 .5) F ( 5 .81 ) 809S ( 3 81 ) T I ME ( 81 ) T ( 1323 )
COMMON KMAX. ASIGMA(5).AKAPPA(S).p(5.e1).H(5t10.81).PS(5lt
C
1 HS ( 5.10) F P R EV ( 5) AMA T R X ( 243.10 ) Pv E CTR ( 243 ) EMA T RX ( 50.50)
2 FVECTR(50) .C(10).V(3.81).NCHANG.ICHANG(5).FCHANG(5).IFLAG
C PRODUCE INITIAL APPROXIMATION
C INTEGRATE NONLINEAR D.E.
IFLAG=l
K=l
L=3
DO 8 1=1.5
L=L+1
8 T ( L ) = A K A P P A ( I ) * E X P ( - O. 5 * ( TAU ( I ) I A S I GM A ( I ) ) * * 2 )
DO 81 1=1.5
L=L+1
81 T(L)=ASIGMA(I)
T(2)=0.0
T(3)=DELTA
CALL INTS(T.10.2.0.0.0.0.0.0.0.0.0.0.0.0)
L=3
DO 82 1=1.5
L=L+1
82 F(].K)=T(L)
C POTENTIAL COMPUTED
C
C
DO 83 J=1.3
V(J.K)=O.O
DO 83 1=1.5
83 V(J.K)=V(J.K) + ADISTN(J.I)*F(].K)
PRINT906
PRINT907.TIME(K).(F(I.K) .1=1.5).(V(J.K).J=1.3)
C INTEGRATE
DO 100 K=2.NMAX
CALL I NTM
L=3
DO 91 1=1. S
L=L+l
91 FII.K)=T(l)
C
C
DO 92 J=1,3
VIJ,K)=O.O
DO 92 1=1,5
92 VIJ,K)=Vlj,K) + ADISTNIJoI)*FII,K)
PRINT907,TIMEIK),IFII,K) ,1=1,5),IVIJ,K) ,J=1,3)
100 CONTINUE
RETURN
906 FORMATIIH08XIHT,5X,5X5HFIIT),5X5HF2IT) ,5X5HF3IT),5X5HF4IT),
1 5X5HF51T),5X,7X5HVIIT),7X5HV2IT),7X5HV3IT)//)
907 FORMATIFI0.4,5X5FI0.4,5X3F12.6)
END
$ IBFTC APPRX2
SUBROUTINE APPRX2
C
C
C
C
C
C
C
C
o I MENS ION PIVOT (50) , I NDEX I 50,2) , I PIVOT I 5(' ) , EM I 50,50) , FV I 50,1 )
COMMON NMAX,DELTA,TAU(5),SIGMAI5),ZKAPPAI5),DCOSI5,3),XCHAMBI5
1 XOB SVR 14,3) , AD I ST N I 3 ,5) , F I 5 ,81) , BOBS I 3 , 8 1 ) , T I ME I 81 ) , T I 1323 )
COMMON KMAX, ASIGMA(5) .AKAPPA(5),PI5,81),HI5,10,81),PSI5),
1 HSI5,10),FPREVI5).AMATRXI243,10),BVECTRI243),EMATRXI50,50),
2 FVECTR(50),CII0).VI3,81),NCHANG,ICHANGI5),FCHANGI5),IFLAG
1=0
APPLY BOUNDARY
SOLVE LINEAR EQUATIONS
COMPUTE NEW APPROXIMATION
SET UP MATRIX, VECTOR
DO 3 J=I,3
DO 3 K = 1,81
1=1+1
DO 1 L=1010
AMA T R X I I , L ) = 0.0
DO 1 M=1.5
AMATRXII.L)=AMATRXII,L) + ADISTNIJ,M)*HIM,L,K)
BVECTRII)=BOBSlj,K)
DO 2 M=105
2 BVECTRII)=BVECTRII) - ADISTNIJ,M)*PIM,K)
3 CONTINUE
DO 6 1=1,10
DO 4 J=lolO
EMATRXII,J)=O.O
DO 4 L=I,243
4 EMATRXII,J)=EMATRXII,J) + AMATRXIL,I)*AMATRXIL,j)
FVECTR I I) =0.0
DO 5 L=1,243
5 FVECTRII)=FVECTRII) + AMATRXIL,I)*BVECTRIL)
6 CONTINUE
C INVERT MATRIX
DO 61 1=1,50
FVII,l)=FVECTRII)
DO 61 J=1,50
APPENDIX SIX 20
(
61 EM(I.JI=EMATRX(I.JI
(ALL MATINV(EM.I0.FV.1.DETERM.PIVOT.INDEx.IPIVOTI
DO 62 1=1.S0
62 FVE(TR ( I 1 = FV ( loll
( NEW APPROXIMATION
(
C
PRINT904
DO 7 1=1.10
7 ((II=FVE(TR(II
DO 8 l=l.S
F(lolJ=((11
ASIGMA (II =(( I+S 1
AKAPPA( I 1 =F (I tl IHXP(0.5* (TAU( I 1 IASIGMA( I 1 1**21
PR I NT 9 05. I AS I GMA ( I 1 AKA P PA ( I 1
PRINT906
K=l
DO 83 J=l03
V(J.KI=O.O
DO 83 1=1.5
83 V(J.KI=V(J.KI + ADISTN(JoII*F(I.KI
PR I NT 9 0 7. T I ME {K 1 { F ( I K 1 1= 1 S I ( V ( J. K 1 J = 1 .31
DO 100 K=2.NMAX
DO 9 1=1. S
F( I .KI=P(! .KI
DO 9 J=lolO
9 F(I.KI=F(I.KI + C(JI*H(I.J.KI
DO 93 J=1.3
V(J.KI=O.O
DO 93 1=1. S
93 V(J.KI=V(J.KI + ADISTN(JoII*F(I.KI
100 PRINT907.TIME(KI.(F(I.KI.I=1.51.(V(J.KI.J=1.31
904
90S
906
RETURN
FORMAT( IHO.8XIHI. 5X 5HSIGMA. 5X 5HKAPPAIII
FORMAT(II0.2FI0.41
FORMA T ( 1 H 0 8 X 1 HT S X. S X S HF 1 ( T I SX SHF 2 ( T 1 .5 X S HF 3 ( T I 5X SHF 4 ( T I
1 5X5HFS(TI.5X,7XSHVl(TI.7XSHV2(TI.7X5HV3(TIIII
907 FORMAT(FI0.4.SX5FI0.4.SX3FI2.61
END
$IBFTC DAUX
(
(
(
(
(
SUBROUTINE DAUX
COMMON NMAX.DELTA.TAU(51.SIGMA(SI.ZKAPPA(SI.D(OS(5.31.XCHAMB(
1 XOBS VR ( 4031 , AD I ST N ( 3 S 1 F ( S , 81 1 BOBS ( 3 B 1 1 or I ME ( 81 1 T ( 1323 I
COMMON KMAX. ASIGMA(SI .AKAPPA(SI.P(5.Bl1.H(5.10.811.PS(51.
1 HS(S,101,FPREV(S),AMATRX(243.10l.BVECTR(2431,EMATRX(SO,501.
2 FVE(TR(501.((101.V(3,811,NCHANG.ICHANG(SI,F(HANG(SI,IFLAG
DIMENSION XF(SI.XSG(51
GO TO (1.2101FLAG
NONLINEAR EQUATIONS
(
(
L=3
M=13
DO 10 1=1,5
L=L+l
M=M+l
10 TCM)=-CTC21-TAUCII)*TCL)/CTCL+51**2)
DO 11 1=1,5
M=M+l
11 TCMIO.O
RETURN
( LINEAR EQS. FOR p, H
(PARTI(ULAR
(
(
2 L=3
DO 3 1=1,5
L=L+l
XFCI)=TCl)
DO 4 1=1,5
L=L+l
4 XSGCI)=TCLI
M=113
DO 5 1=105
M=M+l
TX=- C T C 2) -TAU C I ) I I C AS I GMA C I ) **2 I
5 TCMI=TX*CXFCI) + XSGCI)*C-2.0*FPREV{J)/ASIGMACl)) + 2.0*FPREVCI
DO 6 1=1,5
M=M+l
6 TCM)=O.O
(
(HOMOGENEOUS
(
(
(
DO 13 J=1,10
DO 7 1=1,5
L=L+l
7 XFCI)=TCL)
DO B 1=1,5
L=L+l
8 XSGCII=TCLl
DO 9 1=1,5
M=M+l
TX=-CT(2)-TAUCI))/CASIGMACI)**21
9 TCMI=TX*CXFCI) + XSGCII*C-2.0*FPREVCII/ASIGMACI))
DO 12 1=1,5 '
M=M+l
12 TCM)=O.O
13 (ONTINUE
RETURN
END
INDEX
Abel, N., 60
Adjoint equation, 33
Adorno, D., 108
Approximation in policy space, 19,
124, 154
Averbukh, A. I., 74
Azen, S., 124
Back and forth integration, 85
Bateman-Burgers equation, 122
Beckenbach, E. F., 2, 24, 54, 72,
108, 124, 162
Beckwith, R., 108
Bellman, R., 23, 24, 53, 54, 55, 72,
73, 105, 106, 107, 108, 109, 123,
124, 150, 151, 161, 162
Bentley, D. L., 162
Bihari, 1., 24
Bing, R. H., 72
Blood, J., 6
Bonsall, F. F., 2, 72
Brown, T. A., 6, 105
Buell, J., 6, 106
Calculus of variations, 44, 54
Calogero. F . 24
Caplygin, S. A., 2, 22, 24, 71, 72,
Characteristic values, 66
Chemotherapy, 106, 107
Cherry, 1., 73
Coddington, E., 54
Cohen, E. R., 107
CollatZ, L., 2, 53, 72, 74
Collier, c., 151
Control, 125
Control processes, 54, 73
Control theory, 54
Convergence, 38
Convexity, 34
Convolution equations, 107
Conway, E. D., 124
Cooke, K. L., 106, 162
Culver, J. W., 107
Curtis, P. C., 72
Dewan, E. M., 150
Difference approximations, 73, 1
124
Differential approximation, 75,
106, 158
Differential-difference equations.
106
204
Differential inequalities 57, 71, 74
Dini, D., 21
Dini's theorem, 70, 74
Dreyfus, S., 55, 162
Duality,20
Dynamic programming, 2, 19, 23,
51,53,54,71,79, 103, 124, 153
Electroencepholography, 150
Elliptic equation, 116
Euler equation, 44
Existence and boundedness, 37
Existence and uniqueness, 53
Factorization, 60
Fenchel, W., 2
Fermat's principle, 45
Finite difference methods, 72
First-order linear equation, 15
Fleming, W. H., 124
Forsythe, G. E., 72
Freimer, M., 108
Friedrichs, K. 0.,2, 54
Functional differential equations, 87
Fundamental1emma, 22
Gauss-Seidel method, 118
Glicksberg, I., 54, 73, 108
Gluss, B., 106, 151, 162
Gradient techniques, 114
Gram-Schmidt orthogonalization, 98
Green's function, 34, 54, 69
Gross, 0., 6, 54, 73, 108
Hammer, P. C., 13, 23
Hanson, M. A., 54
Hardy, G. H., 23
Hartman, P., 72, 108
Heron of Alexandria, 12
INDEX
Hopf, E., 124
Hopf-Lax equation, 122
Identification of systems, 106, 159
Ikebe, T., 54
Ill-conditioned linear systems, 109
Ill-conditioning, 100
Inequalities, 54, 72
Inhomogeneous equation, 29
Integro-differential equation, 93
Invariant imbedding, 52, 55, 71, 104,
108, 109, 123
Inverse problem, 130, 146
Jacobian matrix, 77
Jones, G. S., 74
Juncosa, M., 6, 123
Kagiwada, H., 6, 55, 105, 109, 150,
151, 162
Kalaba, R., 23, 24, 25, 54, 55, 73,
105, 106, 107, 108, 109, 123, 124,
150,151,162
Kalman, R. E., 108
Kantorovich, L. V., 22, 23
Kardashov, A. A., 106
Karlin, S., 74
Kato, T., 54
Koepcke, R. W., 108
Kolmogorov, A., 107
Kotkin, B., 6, 73, 105, 106, 107, 123
Krasnoselskii, M. A., 72
Krylov, V. I., 23
Kuhn, H., 2
Kumar, K. S. P., 25
Lagrange, J., 30
Lagrange multipliers, 126
Lanczos, C., 107
Langenhop, C. E., 24
INDEX
Laplace transform, 109
Lavi, A., 151
Lax, P., 124
Levinson, N., 54
Levitt, L. C., 107
Liberman, L. Kh., 107
Linear closure, 107
Linearization, 25
Liouville, J., 23
Lockett, J., 109
Magnethohydrodynamics, 41
Markovian decision processes, 108
Matrix analysis, 53
Maximum operation, 15
Mean-value theorem, 72, 108
Memory, 95
Mesarovic, M. D., 55, 107
Method of steepest descent, 123
Minkowski, A., 2
Mlak, W., 74
Monotone behavior, 57
Monotonicity, 20, 21, 95, 97, 121
Moser, J., 24
Motzkin, T., 72, 73
Multiple-point linear boundary
problems, 105
Multipoint boundary-value prob-
lems,80
Nemytski, V. V., 25
Neurophysiology, 106
Neustadt, L., 54
Neutron transport theory, 162
Newton- Raphson-Kantorovich
approxiination, 19
Newton-Raphson method, 7, 22
Nonlinear oscillations, 150
Nonlinear problems, 25
Nonlinear summability, 116, 123
Numerical integration, 73
Numerical treatment, 72
Optimal design, 125
Orbit determination, 105, 137
Ostrowski, A., 23
Parabolic equations, 112
Parabolic partial differential eq
tion,63
Partial differential equations, 111
Peixoto, M., 2, 72
Periodogram analysis, 139
Perturbation techniques, 106, 107
Peterson, R., 151
Picard algorithm, 40
Picard iteration, 117
Pollack, M., 162
Polya, G., 72, 108
Positivity, 34, 62, 108, 121
Prager, W., 43
Prestrud, M., 55, 108, 109
Property W, 98
Quadratic convergence, 9, 11, 21
Quasilinearization, 2, 13, 36
Radbill, J. R., 25
Radiative transfer, 130, 150
Random walk, 64
Rayleigh, J. W. S., 134
Redheffer, R. M., 2, 72, 74
Reid, W. T., 72
Renewal equation, 91
Renormalization techniques, 107
Riccati differential equation, 104
Riccati equation, 7, 14, 24
Richardson, J. M., 5, 107, 124
Ritt, J. F., 23
Rosenbloom, P., 123
Roth, R., 106, 151, 162
206
Savenko, S. S., 23
Scattering theory, 24
Scher, A., 145
Schlesinger, L., 72, 108
Schrodinger equation, 7
Schwimmer, H. S., 25
Selvester, R., 151
Shortest-route problem, 162
Simultaneous calculation, 81
Smith, M. C., 25
Square roots, 12
Sridhar, R., 6, 25
Stepanov, V. V., 25
Storage, 75, 80, 95
Strauss, J. C., 151
Sturm-Liouville equation, 29, 69
Sturm-Liouville expansion, 66
Successive approximations, 19, 54,
105
Sylvester, J. J., 13, 23
Systems, 76
Temple, G., 25
Thomas-Fermi equation, 54
Tikhomirov, V. M., 107
Titchmarsh, E. c., 72
Tornheim, L., 72
Total positivity, 74
Tucker, A. W., 2
Tukey, J., 23
INDE
Turbulence, 124
Two-point boundary-value problem
27, 105
Ueno, S., 132, 150
Vainberg, M. M., 23
Vala, K., 74
Valiron, G., 72
Van der Pol, B., 134
Van der Pol equation, 134
Variation-diminishing, 73
Variation of parameters, 30
Vashakmadze, T. S., 105
von Karman, T., 25
Walter, W., 162
Wasow, W., 72, 73
Watson, G. N., 23
Weibenson, N., 162
Wengert's numerical method, 151
Whiteside, D. T., 22
Wilkins, J. E., 72
Wing, G. M., 55, 73, 109
Wintner, A., 108
Wronskian, 30, 60, 97
Wu, C. S., 25
Zaidenberg, E. D., 107
Selected RAND Books
BELLMAN, RICHARD. Adaptive Control Processes: A Guided Tour. Princeton,
N.J.: Princeton University Press, 1961.
BELLMAN, RICHARD. Dynamic Programming. Princeton, N.J.: Princeton Uni-
versity Press, 1957.
BELLMAN, RICHARD. Introduction to Matrix Analysis. New York: McGraw-Hill
Book Company, Inc., 1960.
BELLMAN, RICHARD (ed.). Mathematical Optimization Techniques. Berkeley and
Los Angeles: University of California Press, 1963.
BELLMAN, RICHARD, and KENNETH L. COOKE. Differential-Difference Equations.
New York: Academic Press, 1963.
BELLMAN, RICHARD, and STUART E. DREYFUS. Applied Dynamic Programming.
Princeton, N.J.: Princeton University Press, 1962.
BELLMAN, RICHARD E., HARRIET H. KAGIWADA, ROBERT E. KALABA, and
MARCIA C. PRESTRUD. Invariant Imbedding and Time-Dependent Transport
Processes, Modern Analytic and Computational Methods in Science and
Mathematics, Vol. 2. New York: American Elsevier Publishing Company,
Inc., 1964.
BELLMAN, RICHARD E., ROBERT E. KALABA, and MARCIA C. PRESTRUD. Invariant
Imbedding and Radiative Transfer in Slabs of Finite Thickness, Modern
Analytic and Computational Methods in Science and Mathematics, Vol. 1.
New York: American Elsevier Publishing Company, Inc., 1963.
BUCHHEIM, ROBERT W., and the Staff of The RAND Corporation. New Space
Handbook: Astronautics and Its Applications. New York: Vintage Books,
A Division of Random House, Inc., 1963.
DANTZIG, G. B. Linear Programming and Extensions. Princeton, N.J.: Princeton
University Press, 1963.
DRESHER, MELVIN. Games of Strategy: Theory and Applications. Englewood
Cliffs, N.J.: Prentice-Hall, Inc., 1961.
DREYFUS, STUART. Dynamic Programming and the Calculus of Variations.
New York: Academic Press, 1965.
DUBYAGO, A. D. The Determination of Orbits. Translated by R. D. Burke,
G. Gordon, L. N. Rowell, and F. T. Smith. New York: The Macmillan
Company, 1961
EDELEN, DOMINIC G. B. The Structure of Field Space: An Axiomatic Formulation
of Field Physics. Berkeley and Los Angeles: University of California Press,
1962.
FORD, L. R., JR., and D. R. FULKERSON. Flows in Networks. Princeton, N.J.:
Princeton University Press, 1962.
HARRIS, THEODORE E. The Theory of Branching Processes. Berlin, Germany:
Springer-Verlag, 1963; Englewood Cliffs, N.J.: Prentice-Hall, Inc., 1964.
HITCH, CHARLES J., and ROLAND McKEAN. The Economics of Defense in the
Nuclear Age. Cambridge, Mass.: Harvard University Press, 1960.
JUDD, WILLIAM R. (Ed.). State of Stress in the Earth's Crust. New York:
American Elsevier Publishing Company, Inc., 1964.
QUADE, EDWARD S. (ed.). Analysis for Military Decisions. Chicago: Rand
McNally & Company. Amsterdam: North Holland Publishing Company,
1964.
SELIN, IVAN. Detection Theory. Princeton, N.J.: Princeton University Press, 1965.
WILLIAMS, J. D. The Compleat Strategyst: Being a Primer on the Theory of
Games of Strategy. New York: McGraw-Hill Book Company, Inc., 1954.

Quasi Linear Ization

Uploaded by

Document Information

Original Description:

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

Quasi Linear Ization

Uploaded by

Copyright:

Available Formats

Quasilinearization and Nonlinear

You might also like