# i

Lecture Notes on Optimization
Pravin Varaiya
ii
Contents
1 INTRODUCTION 1
2 OPTIMIZATION OVER AN OPEN SET 7
3 Optimization with equality constraints 15
4 Linear Programming 27
5 Nonlinear Programming 49
6 Discrete-time optimal control 75
7 Continuous-time linear optimal control 83
8 Coninuous-time optimal control 95
9 Dynamic programing 121
iii
iv CONTENTS
PREFACE to this edition
Notes on Optimization was published in 1971 as part of the Van Nostrand Reinhold Notes on Sys-
tem Sciences, edited by George L. Turin. Our aim was to publish short, accessible treatments of
graduate-level material in inexpensive books (the price of a book in the series was about ﬁve dol-
lars). The effort was successful for several years. Van Nostrand Reinhold was then purchased by a
conglomerate which cancelled Notes on System Sciences because it was not sufﬁciently proﬁtable.
Books have since become expensive. However, the World Wide Web has again made it possible to
publish cheaply.
Notes on Optimization has been out of print for 20 years. However, several people have been
using it as a text or as a reference in a course. They have urged me to re-publish it. The idea of
making it freely available over the Web was attractive because it reafﬁrmed the original aim. The
only obstacle was to retype the manuscript in LaTex. I thank Kate Klohe for doing just that.
I would appreciate knowing if you ﬁnd any mistakes in the book, or if you have suggestions for
(small) changes that would improve it.
Berkeley, California P.P. Varaiya
September, 1998
v
vi CONTENTS
PREFACE
These Notes were developed for a ten-week course I have taught for the past three years to ﬁrst-year
graduate students of the University of California at Berkeley. My objective has been to present,
in a compact and uniﬁed manner, the main concepts and techniques of mathematical programming
and optimal control to students having diverse technical backgrounds. A reasonable knowledge of
advanced calculus (up to the Implicit Function Theorem), linear algebra (linear independence, basis,
matrix inverse), and linear differential equations (transition matrix, adjoint solution) is sufﬁcient for
The treatment of the topics presented here is deep. Although the coverage is not encyclopedic,
an understanding of this material should enable the reader to follow much of the recent technical
literature on nonlinear programming, (deterministic) optimal control, and mathematical economics.
The examples and exercises given in the text form an integral part of the Notes and most readers will
need to attend to them before continuing further. To facilitate the use of these Notes as a textbook,
I have incurred the cost of some repetition in order to make almost all chapters self-contained.
However, Chapter V must be read before Chapter VI, and Chapter VII before Chapter VIII.
The selection of topics, as well as their presentation, has been inﬂuenced by many of my students
and colleagues, who have read and criticized earlier drafts. I would especially like to acknowledge
the help of Professors M. Athans, A. Cohen, C.A. Desoer, J-P. Jacob, E. Polak, and Mr. M. Ripper. I
also want to thank Mrs. Billie Vrtiak for her marvelous typing in spite of starting from a not terribly
legible handwritten manuscript. Finally, I want to thank Professor G.L. Turin for his encouraging
and patient editorship.
Berkeley, California P.P. Varaiya
November, 1971
vii
viii CONTENTS
Chapter 1
INTRODUCTION
In this chapter, we present our model of the optimal decision-making problem, illustrate decision-
making situations by a few examples, and brieﬂy introduce two more general models which we
cannot discuss further in these Notes.
1.1 The Optimal Decision Problem
These Notes show how to arrive at an optimal decision assuming that complete information is given.
The phrase complete information is given means that the following requirements are met:
1. The set of all permissible decisions is known, and
2. The cost of each decision is known.
When these conditions are satisﬁed, the decisions can be ranked according to whether they incur
greater or lesser cost. An optimal decision is then any decision which incurs the least cost among
the set of permissible decisions.
In order to model a decision-making situation in mathematical terms, certain further requirements
must be satisﬁed, namely,
1. The set of all decisions can be adequately represented as a subset of a vector space with each
vector representing a decision, and
2. The cost corresponding to these decisions is given by a real-valued function.
Some illustrations will help.
Example 1: The Pot Company (Potco) manufacturers a smoking blend called Acapulco Gold.
The blend is made up of tobacco and mary-john leaves. For legal reasons the fraction α of mary-
john in the mixture must satisfy 0 < α <
1
2
. From extensive market research Potco has determined
their expected volume of sales as a function of α and the selling price p. Furthermore, tobacco can
be purchased at a ﬁxed price, whereas the cost of mary-john is a function of the amount purchased.
If Potco wants to maximize its proﬁts, how much mary-john and tobacco should it purchase, and
what price p should it set?
Example 2: Tough University provides “quality” education to undergraduate and graduate stu-
dents. In an agreement signed with Tough’s undergraduates and graduates (TUGs), “quality” is
1
2 CHAPTER 1. INTRODUCTION
deﬁned as follows: every year, each u (undergraduate) must take eight courses, one of which is a
seminar and the rest of which are lecture courses, whereas each g (graduate) must take two seminars
and ﬁve lecture courses. A seminar cannot have more than 20 students and a lecture course cannot
have more than 40 students. The University has a faculty of 1000. The Weary Old Radicals (WORs)
have a contract with the University which stipulates that every junior faculty member (there are 750
of these) shall be required to teach six lecture courses and two seminars each year, whereas every
senior faculty member (there are 250 of these) shall teach three lecture courses and three seminars
each year. The Regents of Touch rate Tough’s President at α points per u and β points per g “pro-
cessed” by the University. Subject to the agreements with the TUGs and WORs how many u’s and
g’s should the President admit to maximize his rating?
Example 3: (See Figure 1.1.) An engineer is asked to construct a road (broken line) connection
point a to point b. The current proﬁle of the ground is given by the solid line. The only requirement
is that the ﬁnal road should not have a slope exceeding 0.001. If it costs \$c per cubic foot to excavate
or ﬁll the ground, how should he design the road to meet the speciﬁcations at minimum cost?
Example 4: Mr. Shell is the manager of an economy which produces one output, wine. There
are two factors of production, capital and labor. If K(t) and L(t) respectively are the capital stock
used and the labor employed at time t, then the rate of output of wine W(t) at time is given by the
production function
W(t) = F(K(t), L(t))
As Manager, Mr. Shell allocates some of the output rate W(t) to the consumption rate C(t), and
the remainder I(t) to investment in capital goods. (Obviously, W, C, I, and K are being measured
in a common currency.) Thus, W(t) = C(t) + I(t) = (1 − s(t))W(t) where s(t) = I(t)/W(t)

.
.
a
b
Figure 1.1: Admissable set of example.
∈ [0, 1] is the fraction of output which is saved and invested. Suppose that the capital stock decays
exponentially with time at a rate δ > 0, so that the net rate of growth of capital is given by the
following equation:
˙
K(t) =
d
dt
K(t) (1.1)
= −δK(t) +s(t)W(t)
= −δK(t) +s(t)F(K(t), L(t)).
The labor force is growing at a constant birth rate of β > 0. Hence,
1.1. THE OPTIMAL DECISION PROBLEM 3
˙
L(t) = βL(t).
(1.2)
Suppose that the production function F exhibits constant returns to scale, i.e., F(λK, λL) =
λF(K, L) for all λ > 0. If we deﬁne the relevant variable in terms of per capita of labor, w =
W/L, c = C/L, k = K/l, and if we let f(k) = F(k, l), then we see that F(K, L)−LF(K/L, 1) =
Lf(k), whence the consumption per capita of labor becomes c(t) = (l −s(t))f(k(t)). Using these
deﬁnitions and equations (1.1) and (1.2) it is easy to see that K(t) satisﬁes the differential equation
(1.3).
˙
k(t) = s(t)f(k(t)) −µk(t)
(1.3)
where µ = (δ +β). The ﬁrst term of the right-hand side in (3) is the increase in the capital-to-labor
ratio due to investment whereas the second terms is the decrease due to depreciation and increase in
the labor force.
Suppose there is a planning horizon time T, and at time 0 Mr. Shell starts with capital-to-labor
ratio k
o
. If “welfare” over the planning period [0, T] is identiﬁed with total consumption

T
0
c(t)dt,
what should Mr. Shell’s savings policy s(t), 0 ≤ t ≤ T, be so as to maximize welfare? What
savings policy maximizes welfare subject to the additional restriction that the capital-to-labor ratio
at time T should be at least k
T
? If future consumption is discounted at rate α > 0 and if time horizon
is ∞, the welfare function becomes

0
e

αt c(t)dt. What is the optimum policy corresponding to
this criterion?
These examples illustrate the kinds of decision-making problems which can be formulated math-
ematically so as to be amenable to solutions by the theory presented in these Notes. We must always
remember that a mathematical formulation is inevitably an abstraction and the gain in precision may
have occurred at a great loss of realism. For instance, Example 2 is caricature (see also a faintly re-
lated but more more elaborate formulation in Bruno [1970]), whereas Example 4 is light-years away
from reality. In the latter case, the value of the mathematical exercise is greater the more insensitive
are the optimum savings policies with respect to the simplifying assumptions of the mathematical
model. (In connection with this example and related models see the critique by Koopmans [1967].)
In the examples above, the set of permissible decisions is represented by the set of all points
in some vector space which satisfy certain constraints. Thus, in the ﬁrst example, a permissible
decision is any two-dimensional vector (α, p) satisfying the constraints 0 < α <
1
2
and 0 <
p. In the second example, any vector (u, g) with u ≥ 0, g ≥ 0, constrained by the number
of faculty and the agreements with the TUGs and WORs is a permissible decision. In the last
example, a permissible decision is any real-valued function s(t), 0 ≤ t ≤ T, constrained by
0 ≤ s(t) ≤ 1. (It is of mathematical but not conceptual interest to note that in this case a decision
is represented by a vector in a function space which is inﬁnite-dimensional.) More concisely then,
these Notes are concerned with optimizing (i.e. maximizing or minimizing) a real-valued function
over a vector space subject to constraints. The constraints themselves are presented in terms of
functional inequalities or equalities.
4 CHAPTER 1. INTRODUCTION
At this point, it is important to realize that the distinction between the function which is to be
optimized and the functions which describe the constraints, although convenient for presenting the
mathematical theory, may be quite artiﬁcial in practice. For instance, suppose we have to choose
the durations of various trafﬁc lights in a section of a city so as to achieve optimum trafﬁc ﬂow.
Let us suppose that we know the transportation needs of all the people in this section. Before we
can begin to suggest a design, we need a criterion to determine what is meant by “optimum trafﬁc
ﬂow.” More abstractly, we need a criterion by which we can compare different decisions, which in
this case are different patterns of trafﬁc-light durations. One way of doing this is to assign as cost to
each decision the total amount of time taken to make all the trips within this section. An alternative
and equally plausible goal may be to minimize the maximum waiting time (that is the total time
spent at stop lights) in each trip. Now it may happen that these two objective functions may be
inconsistent in the sense that they may give rise to different orderings of the permissible decisions.
Indeed, it may be the case that the optimum decision according to the ﬁrst criterion may be lead to
very long waiting times for a few trips, so that this decision is far from optimum according to the
second criterion. We can then redeﬁne the problem as minimizing the ﬁrst cost function (total time
for trips) subject to the constraint that the waiting time for any trip is less than some reasonable
bound (say one minute). In this way, the second goal (minimum waiting time) has been modiﬁed
and reintroduced as a constraint. This interchangeability of goal and constraints also appears at a
deeper level in much of the mathematical theory. We will see that in most of the results the objective
function and the functions describing the constraints are treated in the same manner.
1.2 Some Other Models of Decision Problems
Our model of a single decision-maker with complete information can be generalized along two
very important directions. In the ﬁrst place, the hypothesis of complete information can be relaxed
by allowing that decision-making occurs in an uncertain environment. In the second place, we
can replace the single decision-maker by a group of two or more agents whose collective decision
determines the outcome. Since we cannot study these more general models in these Notes, we
merely point out here some situations where such models arise naturally and give some references.
1.2.1 Optimization under uncertainty.
A person wants to invest \$1,000 in the stock market. He wants to maximize his capital gains, and
at the same time minimize the risk of losing his money. The two objectives are incompatible, since
the stock which is likely to have higher gains is also likely to involve greater risk. The situation
is different from our previous examples in that the outcome (future stock prices) is uncertain. It is
customary to model this uncertainty stochastically. Thus, the investor may assign probability 0.5 to
the event that the price of shares in Glamor company increases by \$100, probability 0.25 that the
price is unchanged, and probability 0.25 that it drops by \$100. A similar model is made for all the
other stocks that the investor is willing to consider, and a decision problem can be formulated as
follows. How should \$1,000 be invested so as to maximize the expected value of the capital gains
subject to the constraint that the probability of losing more than \$100 is less than 0.1?
As another example, consider the design of a controller for a chemical process where the decision
variable are temperature, input rates of various chemicals, etc. Usually there are impurities in the
chemicals and disturbances in the heating process which may be regarded as additional inputs of a
1.2. SOME OTHER MODELS OF DECISION PROBLEMS 5
random nature and modeled as stochastic processes. After this, just as in the case of the portfolio-
selection problem, we can formulate a decision problem in such a way as to take into account these
random disturbances.
If the uncertainties are modelled stochastically as in the example above, then in many cases
the techniques presented in these Notes can be usefully applied to the resulting optimal decision
problem. To do justice to these decision-making situations, however, it is necessary to give great
attention to the various ways in which the uncertainties can be modelled mathematically. We also
need to worry about ﬁnding equivalent but simpler formulations. For instance, it is of great signif-
icance to know that, given appropriate conditions, an optimal decision problem under uncertainty
is equivalent to another optimal decision problem under complete information. (This result, known
as the Certainty-Equivalence principle in economics has been extended and baptized the Separation
Theorem in the control literature. See Wonham [1968].) Unfortunately, to be able to deal with
these models, we need a good background in Statistics and Probability Theory besides the material
presented in these Notes. We can only refer the reader to the extensive literature on Statistical De-
cision Theory (Savage [1954], Blackwell and Girshick [1954]) and on Stochastic Optimal Control
(Meditch [1969], Kushner [1971]).
1.2.2 The case of more than one decision-maker.
Agent Alpha is chasing agent Beta. The place is a large circular ﬁeld. Alpha is driving a fast, heavy
car which does not maneuver easily, whereas Beta is riding a motor scooter, slow but with good
maneuverability. What should Alpha do to get as close to Beta as possible? What should Beta
do to stay out of Alpha’s reach? This situation is fundamentally different from those discussed so
far. Here there are two decision-makers with opposing objectives. Each agent does not know what
the other is planning to do, yet the effectiveness of his decision depends crucially upon the other’s
decision, so that optimality cannot be deﬁned as we did earlier. We need a new concept of rational
(optimal) decision-making. Situations such as these have been studied extensively and an elaborate
structure, known as the Theory of Games, exists which describes and prescribes behavior in these
situations. Although the practical impact of this theory is not great, it has proved to be among the
most fruitful sources of unifying analytical concepts in the social sciences, notably economics and
political science. The best single source for Game Theory is still Luce and Raiffa [1957], whereas
the mathematical content of the theory is concisely displayed in Owen [1968]. The control theorist
will probably be most interested in Isaacs [1965], and Blaquiere, et al., [1969].
The difﬁculty caused by the lack of knowledge of the actions of the other decision-making agents
arises even if all the agents have the same objective, since a particular decision taken by our agent
may be better or worse than another decision depending upon the (unknown) decisions taken by the
other agents. It is of crucial importance to invent schemes to coordinate the actions of the individual
decision-makers in a consistent manner. Although problems involving many decision-makers are
present in any system of large size, the number of results available is pitifully small. (See Mesarovic,
et al., [1970] and Marschak and Radner [1971].) In the author’s opinion, these problems represent
one of the most important and challenging areas of research in decision theory.
6 CHAPTER 1. INTRODUCTION
Chapter 2
OPTIMIZATION OVER AN OPEN
SET
In this chapter we study in detail the ﬁrst example of Chapter 1. We ﬁrst establish some notation
which will be in force throughout these Notes. Then we study our example. This will generalize
to a canonical problem, the properties of whose solution are stated as a theorem. Some additional
properties are mentioned in the last section.
2.1 Notation
2.1.1
All vectors are column vectors, with two consistent exceptions mentioned in 2.1.3 and 2.1.5 below
and some other minor and convenient exceptions in the text. Prime denotes transpose so that if
x ∈ R
n
then x

is the row vector x

= (x
1
, . . . , x
n
), and x = (x
1
, . . . , x
n
)

. Vectors are normally
denoted by lower case letters, the ith component of a vector x ∈ R
n
is denoted x
i
, and different
vectors denoted by the same symbol are distinguished by superscripts as in x
j
and x
k
. 0 denotes
both the zero vector and the real number zero, but no confusion will result.
Thus if x = (x
1
, . . . , x
n
)

and y = (y
1
, . . . , y
n
)

then x

y = x
1
y
1
+ . . . + x
n
y
n
as in ordinary
matrix multiplication. If x ∈ R
n
we deﬁne [x[ = +

x

x.
2.1.2
If x = (x
1
, . . . , x
n
)

and y = (y
1
, . . . , y
n
)

then x ≥ y means x
i
≥ y
i
, i = 1, . . . , n. In particular if
x ∈ R
n
, then x ≥ 0, if x
i
≥ 0, i = 1, . . . , n.
2.1.3
Matrices are normally denoted by capital letters. If A is an m n matrix, then A
j
denotes the jth
column of A, and A
i
denotes the ith row of A. Note that A
i
is a row vector. A
j
i
denotes the entry
of A in the ith row and jth column; this entry is sometimes also denoted by the lower case letter
a
ij
, and then we also write A = ¦a
ij
¦. I denotes the identity matrix; its size will be clear from the
context. If confusion is likely, we write I
n
to denote the n n identity matrix.
7
8 CHAPTER 2. OPTIMIZATION OVER AN OPEN SET
2.1.4
If f : R
n
→ R
m
is a function, its ith component is written f
i
, i = 1, . . . , m. Note that f
i
: R
n
→ R.
Sometimes we describe a function by specifying a rule to calculate f(x) for every x. In this case
we write f : x → f(x). For example, if A is an mn matrix, we can write F : x → Ax to denote
the function f : R
n
→ R
m
whose value at a point x ∈ R
n
is Ax.
2.1.5
If f : R
n
→ Ris a differentiable function, the derivative of f at ˆ x is the row vector ((∂f/∂x
1
)(ˆ x), . . . , (∂f/∂x
n
)(ˆ x)).
This derivative is denoted by (∂f/∂x)(ˆ x) or f
x
(ˆ x) or ∂f/∂x[
x=ˆ x
or f
x
[
x=ˆ x
, and if the argument ˆ x
is clear from the context it may be dropped. The column vector (f
x
(ˆ x))

is also denoted ∇
x
f(ˆ x),
and is called the gradient of f at ˆ x. If f : (x, y) → f(x, y) is a differentiable function from
R
n
R
m
into R, the partial derivative of f with respect to x at the point (ˆ x, ˆ y) is the n-dimensional
row vector f
x
(ˆ x, ˆ y) = (∂f/∂x)(ˆ x, ˆ y) = ((∂f/∂x
1
)(ˆ x, ˆ y), . . . , (∂f/∂x
n
)(ˆ x, ˆ y)), and similarly
f
y
(ˆ x, ˆ y) = (∂f/∂y)(ˆ x, ˆ y) = ((∂f/∂y
1
)(ˆ x, ˆ y), . . . , (∂f/∂y
m
)(ˆ x, ˆ y)). Finally, if f : R
n
→ R
m
is
a differentiable function with components f
1
, . . . , f
m
, then its derivative at ˆ x is the mn matrix
∂f
∂x
(ˆ x) = f
x
ˆ x =

f
1x
(ˆ x)
.
.
.
f
mx
(ˆ x)
¸
¸
¸
=

∂f
1
∂x
1
(ˆ x)
.
.
.
∂fm
∂x
1
(ˆ x)
. . .
. . .
∂f
1
∂xn
(ˆ x)
.
.
.
∂fm
∂xn
(ˆ x)
¸
¸
¸
¸
2.1.6
If f : R
n
→ Ris twice differentiable, its second derivative at ˆ x is the nn matrix (∂
2
f/∂x∂x)(ˆ x) =
f
xx
(ˆ x) where (f
xx
(ˆ x))
j
i
= (∂
2
f/∂x
j
∂x
i
)(ˆ x). Thus, in terms of the notation in Section 2.1.5 above,
f
xx
(ˆ x) = (∂/∂x)(f
x
)

(ˆ x).
2.2 Example
We consider in detail the ﬁrst example of Chapter 1. Deﬁne the following variables and functions:
α = fraction of mary-john in proposed mixture,
p = sale price per pound of mixture,
v = total amount of mixture produced,
f(α, p) = expected sales volume (as determined by market research) of mixture as a function of(α, p).
2.2. EXAMPLE 9
Since it is not proﬁtable to produce more than can be sold we must have:
v = f(α, p),
m = amount (in pounds) of mary-john purchased, and
t = amount (in pounds) of tobacco purchased.
Evidently,
m = αv, and
t = (l −α)v.
Let
P
1
(m) = purchase price of m pounds of mary-john, and
P
2
= purchase price per pound of tobacco.
Then the total cost as a function of α, p is
C(α, p) = P
1
(αf(α, p)) +P
2
(1 −α)f(α, p).
The revenue is
R(α, p) = pf(α, p),
so that the net proﬁt is
N(α, p) = R(α, p) −C(α, p).
The set of admissible decisions is Ω, where Ω = ¦(α, p)[0 < α <
1
2
, 0 < p < ∞¦. Formally, we
have the the following decision problem:
Maximize
subject to
N(α, p),
(α, p) ∈ Ω.
Suppose that (α

, p∗) is an optimal decision, i.e.,

, p

) ∈ Ω
N(α

, p

) ≥ N(α, p)
and
for all (α, p) ∈ Ω.
(2.1)
We are going to establish some properties of (a

, p

). First of all we note that Ω is an open subset
of R
2
. Hence there exits ε > 0 such that
(α, p) ∈ Ω whenever [(α, p) −(α

, p

)[ < ε (2.2)
In turn (2.2) implies that for every vector h = (h
1
, h
2
)

in R
2
there exists η > 0 (η of course
depends on h) such that
((α

, p

) +δ(h
1
, h
2
)) ∈ Ω for 0 ≤ δ ≤ η (2.3)
10 CHAPTER 2. OPTIMIZATION OVER AN OPEN SET
|
.

, p

) +δ(h
1
, h
2
)

α
1
2

δh
h
p
(a

, p

)
Figure 2.1: Admissable set of example.
Combining (2.3) with (2.1) we obtain (2.4):
N(α

, p

) ≥ N(α

+δh
1
, p

+δh
2
) for 0 ≤ δ ≤ η (2.4)
Now we assume that the function N is differentiable so that by Taylor’s theorem
N(α

+δh
1
, p

+δh
2
) =
N(α

, p

)
+δ[
∂N
∂α

, p

)h
1
+
∂N
∂p

, p

)h
2
]
+o(δ),
(2.5)
where

δ
→ 0 as δ → 0. (2.6)
Substitution of (2.5) into (2.4) yields
0 ≥ δ[
∂N
∂α

, p

)h
1
+
∂N
∂p

, p

)h
2
] +o(δ).
Dividing by δ > 0 gives
0 ≥ [
∂N
∂α

, p

)h
1
+
∂N
∂p

, p

)h
2
] +
o(δ)
δ
. (2.7)
Letting δ approach zero in (2.7), and using (2.6) we get
0 ≥ [
∂N
∂α

, p

)h
1
+
∂N
∂p

, p

)h
2
]. (2.8)
Thus, using the facts that N is differentiable, (α

, p

) is optimal, and δ is open, we have concluded
that the inequality (2.9) holds for every vector h ∈ R
2
. Clearly this is possible only if
∂N
∂α

, p

) = 0,
∂N
∂p

, p

) = 0. (2.9)
Before evaluating the usefulness of property (2.8), let us prove a direct generalization.
2.3. THE MAIN RESULT AND ITS CONSEQUENCES 11
2.3 The Main Result and its Consequences
2.3.1 Theorem
.
Let Ω be an open subset of R
n
. Let f: R
n
→ R be a differentiable function. Let x

be an optimal
solution of the following decision-making problem:
Maximize
subject to
f(x)
x ∈ Ω.
(2.10)
Then
∂f
∂x
(x

) = 0. (2.11)
Proof: Since x

∈ Ω and Ω is open, there exists ε > 0 such that
x ∈ Ω whenever [x −x

[ < ε. (2.12)
In turn, (2.12) implies that for every vector h ∈ R
n
there exits η > 0 (η depending on h) such that
(x

+δh) ∈ Ω whenever 0 ≤ δ ≤ η. (2.13)
Since x

is optimal, we must then have
f(x

) ≥ f(x

+δh) whenever 0 ≤ δ ≤ η. (2.14)
Since f is differentiable, by Taylor’s theorem we have
f(x

+δh) = f(x

) +
∂f
∂x
(x

)δh +o(δ), (2.15)
where
o(δ)
δ
→ 0 as δ → 0 (2.16)
Substitution of (2.15) into (2.14) yields
0 ≥ δ
∂f
∂x
(x

)h +o(δ)
and dividing by δ > 0 gives
0 ≥
∂f
∂x
(x

)h +
o(δ)
δ
(2.17)
Letting δ approach zero in (2.17) and taking (2.16) into account, we see that
0 ≥
∂f
∂x
(x

)h, (2.18)
Since the inequality (2.18) must hold for every h ∈ R
n
, we must have
0 =
∂f
∂x
(x

),
and the theorem is proved. ♦
12 CHAPTER 2. OPTIMIZATION OVER AN OPEN SET
Table 2.1
Does there exist At how many points
an optimal deci- in Ω is 2.2.2 Further
Case sion for 2.2.1? satisﬁed? Consequences
1 Yes Exactly one point, x

is the
say x

unique optimal
2 Yes More than one point
3 No None
4 No Exactly one point
5 No More than one point
2.3.2 Consequences.
Let us evaluate the usefulness of (2.11) and its special case (2.18). Equation (2.11) gives us n
equations which must be satisﬁed at any optimal decision x

= (x

1
, . . . , x

n
)

.
These are
∂f
∂x
1
(x

) = 0,
∂f
∂x
2
(x

) = 0, . . . ,
∂f
∂xn
(x

) = 0 (2.19)
Thus, every optimal decision must be a solution of these n simultaneous equations of n variables, so
that the search for an optimal decision from Ω is reduced to searching among the solutions of (2.19).
In practice this may be a very difﬁcult problem since these may be nonlinear equations and it may
be necessary to use a digital computer. However, in these Notes we shall not be overly concerned
with numerical solution techniques (but see 2.4.6 below).
The theorem may also have conceptual signiﬁcance. We return to the example and recall the
N = R − C. Suppose that R and C are differentiable, in which case (2.18) implies that at every
optimal decision (α

, p

)
∂R
∂α

, p

) =
∂C
∂α

, p

),
∂R
∂p

, p

) =
∂C
∂p

, p

),
or, in the language of economic analysis, marginal revenue = marginal cost. We have obtained an
important economic insight.
2.4 Remarks and Extensions
2.4.1 A warning.
Equation (2.11) is only a necessary condition for x

to be optimal. There may exist decisions ˜ x ∈ Ω
such that f
x
(˜ x) = 0 but ˜ x is not optimal. More generally, any one of the ﬁve cases in Table 2.1 may
occur. The diagrams in Figure 2.1 illustrate these cases. In each case Ω = (−1, 1).
Note that in the last three ﬁgures there is no optimal decision since the limit points -1 and +1 are
not in the set of permissible decisions Ω = (−1, 1). In summary, the theorem does not give us any
clues concerning the existence of an optimal decision, and it does not give us sufﬁcient conditions
either.
2.4. REMARKS AND EXTENSIONS 13
Case 1 Case 2 Case 3
Case 5 Case 4
-1
1 -1 1
-1 1
1 -1 -1 1
Figure 2.2: Illustration of 4.1.
2.4.2 Existence.
If the set of permissible decisions Ω is a closed and bounded subset of R
n
, and if f is continuous,
then it follows by the Weierstrass Theorem that there exists an optimal decision. But if Ω is closed
we cannot assert that the derivative of f vanishes at the optimum. Indeed, in the third ﬁgure above,
if Ω = [−1, 1], then +1 is the optimal decision but the derivative is positive at that point.
2.4.3 Local optimum.
We say that x

∈ Ω is a locally optimal decision if there exists ε > 0 such that f(x

) ≥ f(x)
whenever x ∈ Ω and [x

− x[ ≤ ε. It is easy to see that the theorem holds (i.e., 2.11) for local
optima also.
2.4.4 Second-order conditions.
Suppose f is twice-differentiable and let x

∈ Ω be optimal or even locally optimal. Then f
x
(x

) =
0, and by Taylor’s theorem
f(x

+δh) = f(x

) +
1
2
δ
2
h

f
xx
(x

)h +o(δ
2
), (2.20)
where
o(δ
2
)
δ
2
→ 0 as δ → 0. Now for δ > 0 sufﬁciently small f(x

+δh) ≤ f(x

), so that dividing
by δ
2
> 0 yields
0 ≥
1
2
h

f
xx
(x

)h +
o(δ
2
)
δ
2
and letting δ approach zero we conclude that h

f
xx
(x

)h ≤ 0 for all h ∈ R
n
. This means that
f
xx
(x

) is a negative semi-deﬁnite matrix. Thus, if we have a twice differentiable objective function,
we get an additional necessary condition.
2.4.5 Sufﬁciency for local optimal.
Suppose at x

∈ Ω, f
x
(x

) = 0 and f
xx
is strictly negative deﬁnite. But then from the expansion
(2.20) we can conclude that x

is a local optimum.
14 CHAPTER 2. OPTIMIZATION OVER AN OPEN SET
2.4.6 A numerical procedure.
At any point ˜ x ∈ Ω the gradient
x
f(˜ x) is a direction along which f(x) increases, i.e., f(˜ x+ε
x
f(˜ x)) > f(˜ x) for all ε > 0 sufﬁciently small. This observation suggests the following scheme for
ﬁnding a point x

∈ Ω which satisﬁes 2.11. We can formalize the scheme as an algorithm.
Step 1. Pick x
0
∈ Ω. Set i = 0. Go to Step 2.
Step 2. Calculate
x
f(x
i
). If
x
f(x
i
) = 0, stop.
Otherwise let x
i+1
= x
i
+d
i

x
f(x
i
) and go
to Step 3.
Step 3. Set i = i + 1 and return to Step 2.
The step size d
i
can be selected in many ways. For instance, one choice is to take d
i
to be an
optimal decision for the following problem:
Max¦f(x
i
+d
x
f(x
i
))[d > 0, (x
i
+d
x
f(x
i
)) ∈ Ω¦.
This requires a one-dimensional search. Another choice is to let d
i
= d
i−1
if f(x
i
+ d
i−1

x
f(x
i
)) > f(x
i
); otherwise let d
i
= 1/k d
i−1
where k is the smallest positive integer such that
f(x
i
+ 1/k d
i−1

x
f(x
i
)) > f(x
i
). To start the process we let d
−1
> 0 be arbitrary.
Exercise: Let f be continuous differentiable. Let ¦d
i
¦ be produced by either of these choices and
let
¦x
i
¦ be the resulting sequence. Then
1. f(x
i+1
) > f(x
i
) if x
i+1
= x
i
, i
2. if x

∈ Ω is a limit point of the sequence ¦x
i
¦, f
x
(x

) = 0.
For other numerical procedures the reader is referred to Zangwill [1969] or Polak [1971].
Chapter 3
OPTIMIZATION OVER SETS
DEFINED BY EQUALITY
CONSTRAINTS
We ﬁrst study a simple example and examine the properties of an optimal decision. This will
generalize to a canonical problem, and the properties of its optimal decisions are stated in the form
of a theorem. Additional properties are summarized in Section 3 and a numerical scheme is applied
to determine the optimal design of resistive networks.
3.1 Example
We want to ﬁnd the rectangle of maximum area inscribed in an ellipse deﬁned by
f
1
(x, y) =
x
2
a
2
+
y
2
b
2
= α.
(3.1)
The problem can be formalized as follows (see Figure 3.1):
Maximize
subject to
f
0
(x, y)
(x, y) ∈ Ω
= 4xy
= ¦(x, y)[f
1
(x, y) = α¦.
(3.2)
The main difference between problem (3.2) and the decisions studied in the last chapter is that
the set of permissible decisions Ω is not an open set. Hence, if (x

, y

) is an optimal decision we
cannot assert that f
0
(x

, y

) ≥ f
0
(x, y) for all (x, y) in an open set containing (x

, y

). Returning
to problem (3.2), suppose (x

, y

) is an optimal decision. Clearly then either x

= 0 or y

= 0. Let
us suppose y

= 0. Then from ﬁgure 3.1 it is evident that there exist (i)ε > 0, (ii) an open set V
containing (x

, y

), and (iii) a differentiable function g : (x

−ε, x

+ε) → V such that
f
1
(x, y) = α and (x, y) ∈ V iff fy = g(x).
1
(3.3)
In particular this implies that y

= g(x

), and that f
1
(x, g(x)) = α whenever [x − x

[ < ε. Since
1
Note that y

= 0 implies f1y(x

, Y

) = 0, so that this assertion follows from the Implicit Function Theorem. The
assertion is false if y

= 0. In the present case let 0 < ε ≤ a −x

and g(x) = +b[α −(x/a)
2
]
1/2
.
15
16 CHAPTER 3. OPTIMIZATION WITH EQUALITY CONSTRAINTS
) ( |
-
y

g(x)
Tangent plane to
Ω at (x

, y

)
(f
1x
, f
1y
)
V

x

x

Figure 3.1: Illustration of example.
(x

, y

) = (x

, g(x

)) is optimum for (3.2), it follows that x

is an optimal solution for (3.4):
Maximize
subject to
ˆ
f
0
(x) = f
0
(x, g(x))
[x −x

[ < ε.
(3.4)
But the constraint set in (3.4) is an open set (in R
1
) and the objective function
ˆ
f
0
is differentiable,
so that by Theorem 2.3.1,
ˆ
f
0x
(x

) = 0, which we can also express as
f
0x
(x

, y

) +f
0y
(x

, y

)g
x
(x

) = 0 (3.5)
Using the fact that f
1
(x, g(x)) ≡ α for [x −x

[ < ε, we see that
f
1x
(x

, y

) +f
1y
(x

, y

)g
x
(x

) = 0,
and since f
1y
(x

, y

) = 0 we can evaluate g
x
(x

),
g
x
(x

) = −f
−1
1y
f
1x
(x

, y

),
and substitute in (3.5) to obtain the condition (3.6):
f
0x
−f
0y
f
−1
1y
f
1x
= 0 at (x

, y

). (3.6)
Thus an optimal decision (x

, y

) must satisfy the two equations f
1
(x

, y

) = α and (3.6). Solving
these yields
x

=
+

(α/2)
1/2
a , y

=
+

(α/2)
1/2
b.
3.2. GENERAL CASE 17
Evidently there are two optimal decisions, (x

, y

) =
+

(α/2)
1/2
(a, b), and the maximum area is
m(α) = 2αab. (3.7)
The condition (3.6) can be interpreted differently. Deﬁne
λ

= f
0y
f
−1
1y
(x

, y

). (3.8)
Then (3.6) and (3.8) can be rewritten as (3.9):
(f
0x
, f
0y
) = λ

(f
1x
, f
1y
) at (x

, y

) (3.9)
In terms of the gradients of f
0
, f
1
, (3.9) is equivalent to
f
0
(x

, y

) = [f
1
(x

, y

)]λ

, (3.10)
which means that at an optimal decision the gradient of the objective function f
0
is normal to the
plane tangent to the constraint set Ω.
Finally we note that
λ

=
∂m
∂α
. (3.11)
where m(α) = maximum area.
3.2 General Case
3.2.1 Theorem.
Let f
i
: R
n
→ R, i = 0, 1, . . . , m (m < n), be continuously differentiable functions and let x

be
an optimal decision of problem (3.12):
Maximize
subject to
f
0
(x)
f
i
(x) = α
i
, i = 1, . . . , m.
(3.12)
Suppose that at x

the derivatives f
ix
(x

), i = 1, . . . , m, are linearly independent. Then there exists
a vector λ

= (λ

1
, . . . , λ

m
)

such that
f
0x
(x

) = λ

1
f
1x
(x

) +. . . +λ

m
f
mx
(x

) (3.13)
Furthermore, let m(α
1
, . . . , α
m
) be the maximum value of (3.12) as a function of α = (α
1
, . . . , α
m
)

.
Let x

(α) be an optimal decision for (3.12). If x

(α) is a differentiable function of α then m(α) is
a differentiable function of α, and

)

=
∂m
∂α
(3.14)
Proof. Since f
ix
(x

), i = 1, . . . , m, are linearly independent, then by re-labeling the coordinates of
x if necessary, we can assume that the mmmatrix [(∂f
i
/∂x
j
)(x

)], 1 ≤ i, j ≤ m, is nonsingular.
By the Implicit Function Theorem (see Fleming [1965]) it follows that there exist (i) ε > 0, (ii) an
18 CHAPTER 3. OPTIMIZATION WITH EQUALITY CONSTRAINTS
open set V in R
n
containing x

, and (iii) a differentiable function g : U → R
m
, where U =
[(x
m+1
, . . . , x
n
)][ [x
m+
−x

m+
[ < ε, = 1, . . . , n −m], such that
f
i
(x
1
, . . . , x
n
) = α
i
, 1 ≤ i ≤ m, and (x
1
, . . . , x
n
) ∈ V
iff
x
j
= g
j
(x
m+1
, . . . , x
n
), 1 ≤ j ≤ m, and (x
m+1
, . . . , x
n
) ∈ U (3.15)
(see Figure 3.2).
In particular this implies that x

j
= g
j
(x

m+1
, . . . , x

n
), 1 ≤ j ≤ m, and
f
i
(g(x
m+1
, . . . , x
n
), x
m+1
, . . . , x
n
) = α
i
, i = 1, . . . , m. (3.16)
For convenience, let us deﬁne w = (x
1
, . . . , x
m
)

, u = (x
m+1
, . . . , x
n
)

and f = (f
1
, . . . , f
m
)

.
Then, since x

= (w

, u

) = (g(u

), u

) is optimal for (3.12), it follows that u

is an optimal
decision for (3.17):
Maximize
subject to
ˆ
f
0
(u) = f
0
(g(u), u)
u ∈ U.
(3.17)
But U is an open subset of R
n−m
and
ˆ
f
0
is a differentiable function on U (since f
0
and g are
differentiable), so that by Theorem 2.3.1 ,
ˆ
f
0u
(u

) = 0, which we can also express using the chain
rule for derivatives as
ˆ
f
0u
(u

) = f
0w
(x

)g
u
(u

) +f
0u
(x

) = 0. (3.18)
Differentiating (3.16) with respect to u = (x
m+1
, . . . , x
n
)

, we see that
f
w
(x

)g
u
(u

) +f
u
(x

) = 0,
and since the mm matrix f
w
(x

) is nonsingular we can evaluate g
u
(u

),
g
u
(u

) = −[f
w
(x∗)]
−1
f
u
(x

),
and substitute in (3.18) to obtain the condition
−f
0w
f
−1
w
f
u
+f
0u
= 0 at x

= (w

, u

). (3.19)
Next, deﬁne the m-dimensional column vector λ

by

)

= f
0w
f
−1
w
[x

. (3.20)
Then (3.19) and (3.20) can be written as (3.21):
(f
0w
(x

), f
0u
(x

)) = (λ

)

(f
w
(x

), f
u
(x

)). (3.21)
Since x = (w, u), this is the same as
f
0x
(x

) = (λ

)

f
x
(x

) = λ

1
f
1x
(x

) +. . . +λ

m
f
mx
(x

),
3.2. GENERAL CASE 19
.
.
.
.
x
1
, . . . , x
m
x

V
x
m+1
(x
m+1
, . . . , x
n
)
(x

m+1
, . . . , x

n
)
2
U
x
n
Ω =
¦x[f
i
(x) = α
i
¦
i = 1, . . . , m
(x

1
, . . . , x

m
)
g(x
m+1
, . . . , x
n
)
Figure 3.2: Illustration of theorem.
which is equation (3.13).
To prove (3.14), we vary α in a neighborhood of a ﬁxed value, say α. We deﬁne w

(α) =
(x

1
(α), . . . , x

m
(α))

and u

(α) = (x

m+1
(α), . . . , x

(
α))

. By hypothesis, f
w
is nonsingular at
x

(α). Since f(x) and x

(α) are continuously differentiable by hypothesis, it follows that f
w
is
nonsingular at x

(α) in a neighborhood of α, say N. We have the equation
f(w

(α), u

(α)) = α, (3.22)
−f
0w
f
−1
w
f
u
+f
0u
= 0 at (w

(α), u

(α)), (3.23)
for α ∈ N. Also, m(α) = f
0
(x

(α)), so that
m
α
= f
0w
w

α
+f
0u
u

α
(3.24)
Differentiating (3.22) with respect to α gives
f
w
w

α
+f
u
u

α
= I,
so that
w

α
+f
−1
w
f
u
u

α
= f
−1
w
,
20 CHAPTER 3. OPTIMIZATION WITH EQUALITY CONSTRAINTS
and multiplying on the left by f
0w
gives
f
0w
w

α
+f
0w
f
−1
w
f
u
u

α
= f
0w
f
−1
w
.
Using (3.23), this equation can be rewritten as
f
0w
w

α
+f
0u
u

α
= f
0w
f
−1
w
. (3.25)
In (3.25), if we substitute from (3.20) and (3.24), we obtain (3.14) and the theorem is proved. ♦
3.2.2 Geometric interpretation.
The equality constraints of the problem in 3.12 deﬁne a n −m dimensional surface
Ω = ¦x[f
i
(x) = α
i
, i = 1, . . . , m¦.
The hypothesis of linear independence of ¦f
ix
(x

)[1 ≤ i ≤ m¦ guarantees that the tangent plane
through Ω at x

is described by
¦h[f
ix
(x

)h = 0 , i = 1, . . . , m¦, (3.26)
so that the set of (column vectors orthogonal to this tangent surface is
¦λ
1

x
f
1
(x

) +. . . +λ
m

x
f
m
(x

)[λ
i
∈ R, i = 1, . . . , m¦.
Condition (3.13) is therefore equivalent to saying that at an optimal decision x

objective function
x
f
0
(x

) is normal to the tangent surface (3.12).
3.2.3 Algebraic interpretation.
Let us again deﬁne w = (x
1
, . . . , x
m
)

and u = (x
m+1
, . . . , x
n
)

. Suppose that f
w
(˜ x) is nonsin-
gular at some point ˜ x = ( ˜ w, ˜ u) in Ω which is not necessarily optimal. Then the Implicit Function
Theorem enables us to solve, in a neighborhood of ˜ x, the mequations f(w, u) = α. u can then vary
arbitrarily in a neighborhood of ˜ u. As u varies, w must change according to w = g(u) (in order to
maintain f(w, u) = α), and the objective function changes according to
ˆ
f
0
(u) = f
0
(g(u), u). The
derivative of
ˆ
f
0
at ˜ u is
ˆ
f
0u
(˜ u) = f
0w
g
u
+f
0u˜ x
= −
˜
λ

f
u
(˜ x) +f
0u
(˜ x),
where
˜
λ

= f
0w
f
−1
w˜ x
, (3.27)
Therefore, the direction of steepest increase of
ˆ
f
0
at ˜ u is

u
ˆ
f
0
(˜ u) = −f

u
(˜ x)
˜
λ +f

Ou
(˜ x) , (3.28)
and if ˜ u is optimal,
u
ˆ
f
0
(˜ u) = 0 which, together with (3.27) is equation (3.13). We shall use (3.27)
and (3.28) in the last section.
3.3. REMARKS AND EXTENSIONS 21
3.3 Remarks and Extensions
3.3.1 The condition of linear independence.
The necessary condition (3.13) need not hold if the derivatives f
ix
(x

), 1 ≤ i ≤ m, are not linearly
independent. This can be checked in the following example
Minimize
subject to sin(x
2
1
+x
2
2
)
π
2
(x
2
1
+x
2
2
) = 1.
(3.29)
3.3.2 An alternative condition.
Keeping the notation of Theorem 3.2.1, deﬁne the Lagrangian function L : R
n+m
→ R by L :
(x, λ) → f
0
(x) −
¸
m
i=1
λ
i
f
i
(x). The following is a reformulation of 3.12, and its proof is left as
an exercise.
Let x

be optimal for (3.12), and suppose that f
ix
(x

), 1 ≤ i ≤ m, are linearly independent.
Then there exists λ

∈ R
m
such that (x

, λ

) is a stationary point of L, i.e., L
x
(x

, λ

) = 0 and
L
λ
(x

, λ

) = 0.
3.3.3 Second-order conditions.
Since we can convert the problem (3.12) into a problem of maximizing
ˆ
f
0
over an open set, all
the comments of Section 2.4 will apply to the function
ˆ
f
0
. However, it is useful to translate these
remarks in terms of the original function f
0
and f. This is possible because the function g is
uniquely speciﬁed by (3.16) in a neighborhood of x

. Furthermore, if f is twice differentiable, so
is g (see Fleming [1965]). It follows that if the functions f
i
, 0 ≤ i ≤ m, are twice continuously
differentiable, then so is
ˆ
f
0
, and a necessary condition for x

to be optimal for (3.12) and (3.13) and
the condition that the (n − m) (n − m) matrix
ˆ
f
0uu
(u

) is negative semi-deﬁnite. Furthermore,
if this matrix is negative deﬁnite then x

is a local optimum. the following exercise expresses
f
ˆ
f
0uu
(u

) in terms of derivatives of the functions f
i
.
Exercise: Show that
ˆ
f
0uu
(u

) = [g

u
.
.
.I]
¸
L
ww
L
uw
L
wu
L
uu

g
u
. . .
I
¸
¸

(w

, u

)
where
g
u
(u

) = −[f
w
(x

)]
−1
f
u
(x

), L(x) = f
0
(x) −
m
¸
i=1
λ

i
f
i
(x).
22 CHAPTER 3. OPTIMIZATION WITH EQUALITY CONSTRAINTS
3.3.4 A numerical procedure.
We assume that the derivatives f
ix
(x), 1 ≤ i ≤ m, are linearly independent for all x. Then the
following algorithm is a straightforward adaptation of the procedure in Section 2.4.6.
Step 1. Find x
0
arbitrary so that f
i
(x
0
) = α
i
, 1 ≤ i ≤ m. Set k = 0 and go to Step 2.
Step 2. Find a partition x = (w, u)
2
of the variables such that f
w
(x
k
) is nonsingular. Calculate λ
k
by (λ
k
)

= f
0w
f
−1
w(xk)
, and
ˆ
f
k
0
(u
k
) = −f

u
(x
k

k
+f

0u
(x
k
). If
ˆ
f
k
0
(u
k
) = 0, stop. Otherwise
go to Step 3.
Step 3. Set ˜ u
k
= u
k
+d
k

ˆ
f
k
0
(u
k
). Find ˜ w
k
such that f
i
( ˜ w
k
, ˜ u
k
) = 0, 1 ≤ i ≤ m. Set
x
k+1
= ( ˜ w
k
, ˜ u
k
), set k = k + 1, and return to Step 2.
Remarks. As before, the step sizes d
k
> 0 can be selected various ways. The practical applicability
of the algorithm depends upon two crucial factors: the ease with which we can ﬁnd a partition
x = (w, u) so that f
w
(x
k
) is nonsingular, thus enabling us to calculate λ
k
; and the ease with which
we can ﬁnd ˜ w
k
so that f( ˜ w
k
, ˜ u
k
) = α. In the next section we apply this algorithm to a practical
problem where these two steps can be carried out without too much difﬁculty.
3.3.5 Design of resistive networks.
Consider a network N with n + 1 nodes and b branches. We choose one of the nodes as datum
and denote by e = (e
1
, . . . , e
n
)

the vector of node-to-datum voltages. Orient the network graph
and let v = (v
1
, . . . , v
b
)

and j = (j
1
, . . . , j
b
)

respectively, denote the vectors of branch voltages
and branch currents. Let A be the n b reduced incidence matrix of the network graph. Then the
Kirchhoff current and voltage laws respectively yield the equations
Aj = 0 and A

e = v (3.30)
Next we suppose that each branch k contains a (possibly nonlinear)resistive element with the form
shown in Figure 3.3, so that
j
k
−j
sk
= g
k
(v
rk
) = g
k
(v
k
−v
sk
), 1 ≤ k ≤ b, (3.31)
where v
rk
is the voltage across the resistor. Here j
sk
, v
sk
are the source current and voltage in the
kth branch, and g
k
is the characteristic of the resistor. Using the obvious vector notation j
s
∈ R
b
,
v
s
∈ R
b
for the sources, v
r
∈ R
b
for the resistor voltages, and g = (g
1
, . . . , g
b
)

, we can rewrite
(3.30) as (3.31):
j −j
s
= g(v −v
s
) = g(v
r
). (3.32)
Although (3.30) implies that the current (j
k
−j
s
k) through the kth resistor depends only on the
voltage v
rk
= (v
k
−v
sk
) across itself, no essential simpliﬁcation is achieved. Hence, in (3.31) we
shall assume that g
k
is a function of v
r
. This allows us to include coupled resistors and voltage-
controlled current sources. Furthermore, let us suppose that there are design parameters p =
(p
1
, . . . , p

)

which are under our control, so that (3.31) is replaced by (3.32):
j −j
x
= g(v
r
, p) = g(v−v
s
, p). (3.33)
2
This is just a notational convenience. The w variable may consist of any m components of x.
3.3. REMARKS AND EXTENSIONS 23
o
-
+ -
+ -
+
o
j
sk
v
rk
v
sk
j
k
−j
sk
j
k
v
k
Figure 3.3: The kth branch.
If we combine (3.29) and (3.32) we obtain (3.33):
Ag(A

e −v
s
, p) = i
s
, (3.34)
where we have deﬁned i
s
= A
js
. The network design problem can then be stated as ﬁnding p, v
s
, i
s
so as to minimize some speciﬁed function f
0
(e, p, v
s
, i
s
). Formally, we have the optimization prob-
lem (3.34):
Minimize
subject to
f
0
(e, p, v
s
, i
s
)
Ag(A

e −v
s
, p) −i
s
= 0.
(3.35)
We shall apply the algorithm 3.3.4 to this problem. To do this we make the following assumption.
Assumption: (a) f
0
is differentiable. (b) g is differentiable and the nn matrix A(∂g/∂v)(v, p)A

is nonsingular for all v ∈ R
b
, p ∈ R

. (c) The network N described by (3.33) is determinate i.e.,
for every value of (p, v
s
, i
s
) there is a unique e = E(p, v
s
, i
s
) satisfying (3.33).
In terms of the notation of 3.3.4, if we let x = (e, p, v
s
, i
s
), then assumption (b) allows us to
identify w = e, and u = (p, v
s
, i
s
). Also let f(x) = f(e, p, v
s
, i
s
) = Ag(A

e−v
s
, p) −i
s
. Now the
crucial part in the algorithm is to obtain λ
k
at some point x
k
. To this end let ˜ x = (˜ e, ˜ p, ˜ v
s
,
˜
i
s
) be a
ﬁxed point. Then the corresponding λ =
˜
λ is given by (see (3.27))
˜
λ

= f
0w
(˜ x)f
−1
w
(˜ x) = f
0e
(˜ x)f
−1
e
(˜ x). (3.36)
From the deﬁnition of f we have
f
e
(˜ x) = AG(˜ v
r
, ˜ p)A

,
where ˜ v
r
= A

˜ e − ˜ v
s
, and G(˜ v
r
, ˜ p) = (∂g/∂v
r
)(˜ v
r
, ˜ p). Therefore,
˜
λ is the solution (unique by
assumption (b)) of the following linear equation:
AG

(˜ v
r
, ˜ p)A

˜
λ = f

0e
(˜ x). (3.37)
Now (3.36) has the following extremely interesting physical interpretation. If we compare (3.33)
with (3.36) we see immediately that
˜
λ is the node-to-datum response voltages of a linear network
N(˜ v
r
, ˜ p) driven by the current sources f

0e
(˜ x). Furthermore, this network has the same graph as
the original network (since they have the same incidence matrix); moreover, its branch admittance
matrix, G

(˜ v
r
, ˜ p), is the transpose of the incremental branch admittance matrix (evaluated at (˜ v
r
, ˜ p))
of the original network N. For this reason, N(˜ v
r
, ˜ p) is called the adjoint network (of N) at (˜ v
r
, ˜ p).
24 CHAPTER 3. OPTIMIZATION WITH EQUALITY CONSTRAINTS
Once we have obtained
˜
λ we can obtain
u
ˆ
f
0
(˜ u) using (3.28). Elementary calculations yield
(3.37):

u
ˆ
f
0
(˜ u) =

ˆ
f

0p
(˜ u)
ˆ
f

0vs
(˜ u)
ˆ
f

0is
(˜ u)
¸
¸
¸ =

[
∂g
∂p
(˜ v
r
, ˜ p)]

A

G

(˜ v
r
, ˜ p)A

−I
¸
¸ ˜
λ +

f

0p
(˜ x)
f

0vs
(˜ x)
f

0is
(˜ x)
¸
¸
(3.38)
We can now state the algorithm.
Step 1. Select u
0
= (p
0
, v
0
s
, i
0
s
) arbitrary. Solve (3.33) to obtain e
0
= E(p
0
, v
0
s
, i
0
s
). Let k = 0 and
go to Step 2.
Step 2. Calculate v
k
r
= A

e
k
−v
k
s
. calculate f

0e
(x
k
). Calculate the node-to-datum response λ
k
of
k
r
, p
k
) driven by the current source f

0e
(x
k
). Calculate
u
ˆ
f
0
(u
k
) from
(3.37). If this gradient is zero, stop. Otherwise go to Step 3.
Step 3. Let u
k+1
= (p
k+1
, v
k+1
s
, i
k+1
s
) = u
k
−d
k

u
ˆ
f
0
(u
k
), where d
k
> 0 is a predetermined
step size.
3
Solve (3.33) to obtain e
k+1
= (Ep
k+1
, v
k+1
s
, i
k+1
s
Remark 1. Each iteration from u
k
to u
k+1
requires one linear network analysis step (the
computation of λ
k
in Step 2), and one nonlinear network analysis step (the computation of e
k+1
in
step 3). This latter step may be very complex.
Remark 2. In practice we can control only some of the components of v
s
and i
s
, the rest being
ﬁxed. The only change this requires in the algorithm is that in Step 3 we set
p
k+1
= p
k
−d
k
ˆ
f

0p
(u
k
) just as before, where as v
k+1
sj
= v
k
sj
−d
k
(∂
ˆ
f
0
/∂v
sj
)(u
k
) and
i
k+1
sm
= i
k
sm
−d
k
(∂
ˆ
f
0
/∂i
sm
)(u
k
) with j and m ranging only over the controllable components and
the rest of the components equal to their speciﬁed values.
Remark 3. The interpretation of λ as the response of the adjoint network has been exploited for
particular function f
0
in a series of papers (director and Rohrer [1969a], [1969b], [1969c]). Their
derivation of the adjoint network does not appear as transparent as the one given here. Although
we have used the incidence matrix A to obtain our network equation (3.33), one can use a more
general cutset matrix. Similarly, more general representations of the resistive elements may be
employed. In every case the “adjoint” network arises from a network interpretation of (3.27),
[f
w
(˜ x)]

˜
λ = f
0w
(˜ x),
with the transpose of the matrix giving rise to the adjective “adjoint.”
Exercise: [DC biasing of transistor circuits (see Dowell and Rohrer [1971]).] Let N be a transistor
circuit, and let (3.33) model the dc behavior of this circuit. Suppose that i
s
is ﬁxed, v
sj
for j ∈ J
are variable, and v
sj
for j / ∈ J are ﬁxed. For each choice of v
sj
, j ∈ J, we obtain the vector e and
hence the branch voltage vector v = A

e. Some of the components v
t
, t ∈ T, will correspond to
bias voltages for the transistors in the network, and we wish to choose v
sj
, j ∈ J, so that v
t
is as
close as possible to a desired bias voltage v
d
t
, t ∈ T. If we choose nonnegative numbers α
t
, with
relative magnitudes reﬂecting the importance of the different transistors then we can formulate the
criterion
3
k
−d
k
u
ˆ
f0(u
k
). Remember we are minimizing f0, which is equivalent to
maximizing (−f0).
3.3. REMARKS AND EXTENSIONS 25
f
0
(e) =
¸
t∈T
α
t
[v
t
−v
d
t
[
2
.
(i) Specialize the algorithm above for this particular case.
(ii) How do the formulas change if the network equations are written using an arbitrary cutset matrix
26 CHAPTER 3. OPTIMIZATION WITH EQUALITY CONSTRAINTS
Chapter 4
OPTIMIZATION OVER SETS
DEFINED BY INEQUALITY
CONSTRAINTS: LINEAR
PROGRAMMING
In the ﬁrst section we study in detail Example 2 of Chapter I, and then we deﬁne the general linear
programming problem. In the second section we present the duality theory for linear program-
ming and use it to obtain some sensitivity results. In Section 3 we present the Simplex algorithm
which is the main procedure used to solve linear programming problems. In section 4 we apply
the results of Sections 2 and 3 to study the linear programming theory of competitive economy.
cessible treatment of the material presented in this chapter see the companion volume in this Series
(Sakarovitch [1971]).
4.1 The Linear Programming Problem
4.1.1 Example.
Recall Example 2 of Chapter I. Let g and u respectively be the number of graduate and undergradu-
ate students admitted. Then the number of seminars demanded per year is
2g+u
20
, and the number of
lecture courses demanded per year is
5g+7u
40
. On the supply side of our accounting, the faculty can
offer 2(750) + 3(250) = 2250 seminars and 6(750) + 3(250) = 5250 lecture courses. Because of
his contractual agreements, the President must satisfy
2g+u
20
≤ 2250 or 2g +u ≤ 45, 000
and
5g+7u
40
≤ 5250 or 5g + 7u ≤ 210, 000 .
27
28 CHAPTER 4. LINEAR PROGRAMMING
Since negative g or u is meaningless, there are also the constraints g ≥ 0, u ≥ 0. Formally then the
President faces the following decision problem:
Maximize αg +βu
subject to 2g +u ≤ 45, 000
5g + 7u ≤ 210, 000
g ≥ 0, u ≥ 0 .
(4.1)
It is convenient to use a more general notation. So let x = (g, u)

, c = (α, β)

, b = (45000, 210000, 0, 0)

and let A be the 42 matrix
A =

2
5
−1
0
1
7
0
−1
¸
¸
¸
¸
.
Then (4.1) can be rewritten as (4.2)
1
Maximize c

x
subject to Ax ≤ b .
(4.2)
Let A
i
, 1 ≤ i ≤ 4, denote the rows of A. Then the set Ω of all vectors x which satisfy the constraints
in (4.2) is given by Ω = ¦x[A
i
x ≤ b
i
, 1 ≤ i ≤ 4¦ and is the polygon OPQR in Figure 4.1.
For each choice x, the President receives the payoff c

x. Therefore, the surface of constant payoff
k say, is the hyperplane π(k) = ¦x[c

x = k¦. These hyperplanes for different values of k are
parallel to one another since they have the same normal c. Furthermore, as k increases π(k) moves
in the direction c. (Obviously we are assuming in this discussion that c = 0.) Evidently an optimal
decision is any point x

∈ Ω which lies on a hyperplane π(k) which is farthest along the direction
c. We can rephrase this by saying that x∗ ∈ Ω is an optimal decision if and only if the plane π

through x

does not intersect the interior of Ω, and futhermore at x

the direction c points away
from Ω. From this condition we can immediately draw two very important conclusions: (i) at least
one of the vertices of Ω is an optimal decision, and (ii) x

yields a higher payoff than all points
in the cone K

consisting of all rays starting at x

and passing through Ω, since K

lies “below”
π

. The ﬁrst conclusion is the foundation of the powerful Simplex algorithm which we present in
Section 3. Here we pursue consequences of the second conclusion. For the situation depicted in
Figure 4.1 we can see that x

= Q is an optimal decision and the cone K

is shown in Figure 4.2.
Now x

satisﬁes A
x
x

= b
1
, A
2
x

= b
2
, and A
3
x

< b
3
, A
4
x

< b
4
, so that K

is given by
K

= ¦x

+h[A
1
h ≤ 0 , A
2
h ≤ 0¦ .
Since c

x

≥ c

y for all y ∈ K

we conclude that
c

h ≤ 0 for all h such that A
1
h ≤ 0, A
2
h ≤ 0 . (4.3)
We pause to formulate the generalization of (4.3) as an exercise.
1
Recall the notation introduced in 1.1.2, so that x ≤ y means xi ≤ yi for all i.
4.1. THE LINEAR PROGRAMMING PROBLEM 29
,
-
-
-
-
-
-
-
-
-
-
x
2
π(k) = ¦x[c

x = k¦
π

Q = x

direction of
increasing
payoff k
¦x[A
2
x = b
2
¦
x
1
¦x[A
1
x = b
1
¦
R
A
4
O A
3
A
1
⊥ QR
c ⊥ π

A
2
⊥ PQ
P
Figure 4.1: Ω = OPQR.
Exercise 1: Let A
i
, 1 ≤ i ≤ k, be n-dimensional row vectors. Let c ∈ R
n
, and let b
i
, 1 ≤ i ≤ k,
be real numbers. Consider the problem
Maximize c

x
subject to A
i
x ≤ b
i
, 1 ≤ i ≤ k .
For any x satisfying the constraints, let I(x) ⊂ ¦1, . . . , n¦ be such that A
i
(x) = b
i
, i ∈ I(x), A
i
x <
b
i
, i / ∈ I(x). Suppose x

satisﬁes the constraints. Show that x

is optimal if an only if
c

h ≤ 0 for all h such that A
i
h ≤ 0 , i ∈ I(x

).
Returning to our problem, it is clear that (4.3) is satisﬁed as long as c lies between A
1
and A
2
.
Mathematically this means that (4.3) is satisﬁed if and only if there exist λ

1
≥ 0, λ

2
≥ 0 such that
2
c

= λ

1
, A
1

2
A
2
. (4.4)
As c varies, the optimal decision will change. We can see from our analysis that the situation is as
follows (see Figure 4.1):
2
Although this statement is intuitively obvious, its generalization to n dimensions is a deep theorem known as Farkas’
lemma (see Section 2).
30 CHAPTER 4. LINEAR PROGRAMMING
P
x

= Q
K

π

R
A
4
O
A
3
A
2
c
A
1
Figure 4.2: K

is the cone generated by Ω at x

.
1. x

= Qis optimal iff c lies between A
1
and A
2
iff c

= λ

1
A
1

2
A
2
for some λ

1
≥ 0, λ

2

0,
2. x

∈ QP is optimal iff c lies along A
2
iff c

= λ

2
A
2
for some λ

2
≥ 0,
3. x

= P is optimal iff c lies between A
3
and A
2
iff c

= λ

2
A
2

3
A
3
for some λ

2
≥ 0, λ

3

0, etc.
These statements can be made in a more elegant way as follows:
x

∈ Ω is optimal iff there exists λ

i
≥ 0 , 1 ≤ i ≤ 4, such that
(a) c

=
4
¸
i=1
λ

i
a
i
, (b) if A
i
x

< b
i
then λ

i
= 0 . (4.5)
For purposes of application it is useful to separate those constraints which are of the form x
i
≥ 0,
from the rest, and to reformulate (4.5) accordingly We leave this as an exercise.
Exercise 2: Show that (4.5) is equivalent to (4.6), below. (Here A
i
= (a
i1
, a
i2
).) x

∈ Ω is optimal
iff there exist λ

1
≥ 0 , λ

2
≥ 0 such that
(a) c
i
≤ λ

1
a
1i

2
a
2i
, i = 1, 2,
(b) if a
j1
x

1
+a
j2
x

2
< b
j
then x

j
= 0, j = 1, 2.
(c) if c
i
< λ

1i

2
a
2i
then x

i
= 0, i = 1, 2.
(4.6)
4.1. THE LINEAR PROGRAMMING PROBLEM 31
4.1.2 Problem formulation.
A linear programming problem (or LP in brief) is any decision problem of the form 4.7.
Maximize c
1
x
1
+c
2
x
2
+. . . +c
n
x
n
subject to
a
il
x
1
+a
i2
x
2
+. . . +a
in
x
n
≤ b
i
, l ≤ i ≤ k ,
a
il
x
1
+. . . . . . . . . +a
in
x
n
≥ b
i
, k + 1 ≤ i ≤ ,
a
il
x
1
+. . . . . . . . . +a
in
x
n
= b
i
, + 1 ≤ i ≤ m ,
and
x
j
≥ 0 , 1 ≤ j ≤ p ,
x
j
≥ 0 , p + 1 ≤ j ≤ q;
x
j
arbitary , q + 1 ≤ j ≤ n ,
(4.7)
where the c
j
, a
ij
, b
i
are ﬁxed real numbers.
There are two important special cases:
Case I: (4.7) is of the form (4.8):
Maximize
n
¸
j=1
c
j
x
j
subject to
n
¸
j=1
a
ij
x
j
≤ b
i
,
x
j
≥ 0 ,
1 ≤ i ≤ m ,
1 ≤ j ≤ n
(4.8)
Case II: (4.7) is of the form (4.9):
Maximize
n
¸
j=1
c
j
x
j
subject to
n
¸
j=1
a
ij
x
j
= b
i
,
x
j
≥ 0 ,
1 ≤ i ≤ m ,
1 ≤ j ≤ n .
(4.9)
Although (4.7) appears to be more general than (4.8) and (4.9), such is not the case.
Proposition: Every LP of the form (4.7) can be transformed into an equivalent LP of the form (4.8).
Proof.
Step 1: Replace each inequality constraint
¸
a
ij
x
j
≥ b
i
by
¸
(−a
ij
)x
j
≤ (−b
i
).
Step 2: Replace each equality constraint
¸
a
ij
x
j
= b
i
by two inequality constraints:
¸
a
ij
x
j
≤ b
i
,
¸
(−a
ij
)x
j
≤ (−b
i
).
Step 3: Replace each variable x
j
which is constrained x
j
≤ 0 by a variable y
j
= −x
j
constrained
y
j
≥ 0 and then replace a
ij
x
j
by (−a
ij
)y
j
for every i and c
j
x
j
by (−c
j
)y
j
.
32 CHAPTER 4. LINEAR PROGRAMMING
Step 4: Replace each variable x
j
which is not constrained in sign by a pair of variables
y
j
−z
j
= x
j
constrained y
j
≥ 0, z
j
≥ 0 and then replace a
ij
x
j
by a
ij
y
j
+ (−a
ij
)z
j
for every i and
c
j
x
j
by c
j
y
j
+ (−c
j
)z
j
. Evidently the resulting LP has the form (4.8) and is equivalent to the
original one. ♦
Proposition: Every LP of the form (4.7) can be transformed into an equivalent LP of the from (4.9)
Proof.
Step 1: Replace each inequality constraint
¸
a
ij
x
j
≤ b
i
by the equality constraint
¸
a
ij
x
j
+y
i
= b
i
where y
i
is an additional variable constrained y
i
≥ 0.
Step 2: Replace each inequality constraint
¸
a
ij
x
j
≥ b
i
by the equality constraint
¸
a
ij
x
j
−y
i
= b
i
where y
i
is an additional variable constrained by y
i
≥ 0. (The new variables
added in these steps are called slack variables.)
Step 3, Step 4: Repeat these steps from the previous proposition. Evidently the new LP has the
form (4.9) and is equivalent to the original one. ♦
4.2 Qualitative Theory of Linear Programming
4.2.1 Main results.
We begin by quoting a fundamental result. For a proof the reader is referred to (Mangasarian
[1969]).
Farkas’ Lemma. Let A
i
, 1 ≤ i ≤ k, be n-dimensional row vectors. Let c ∈ R
n
be a column vector.
The following statements are equivalent:
(i) for all x ∈ R
n
, A
i
x ≤ 0 for 1 ≤ i ≤ k implies c

x ≤ 0,
(ii) there exists λ
1
≥ 0, . . . , λ
k
≥ 0 such that c

=
k
¸
i=1
λ
i
A
i
.
An algebraic version of this result is sometimes more convenient.
Farkas’ Lemma (algebraic version). Let Abe a kn matrix. Let c ∈ R
n
. The following statements
are equivalent.
(i) for all x ∈ R
n
, Ax ≤ 0 implies c

x ≤ 0,
(ii) there exists λ ≥ 0, λ ∈ R
k
, such that A

λ = c.
Using this result it is possible to derive the main results following the intuitive reasoning of (4.1).
We leave this development as two exercises and follow a more elegant but less intuitive approach.
Exercise 1: With the same hypothesis and notation of Exercise 1 in 4.1, use the ﬁrst version of
Farkas

lemma to show that there exist λ

i
≥ 0 for i ∈ I(x

) such that
¸
i∈I(x

)
λ

i
A
i
= c

.
Exercise 2: Let x

satisfy the constraints for problem (4.17). Use the previous exercise to show
that x

is optimal iff there exist λ

1
≥ 0, . . . , λ

m
≥ 0 such that
(a) c
j

m
¸
i=1
λ

i
a
ij
, 1 ≤ j ≤ n
(b) if
n
¸
j=1
a
ij
x

j
< b
i
then λ

i
= 0 , 1 ≤ i ≤ m (c) if
m
¸
i=1
λ

i
a
ij
> c
j
then x

j
= 0 , 1 ≤ j ≤ m.
In the remaining discussion, c ∈ R
n
, b ∈
n
are ﬁxed vectors, and A = ¦a
ij
¦ is a ﬁxed m n
matrix, whereas x ∈ R
n
and λ ∈ R
m
will be variable. Consider the pair of LPs (4.10) and (4.11)
4.2. QUALITATIVE THEORY OF LINEAR PROGRAMMING 33
below. (4.10) is called the primal problem and (4.11) is called the dual problem.
Maximize
subject to
c
1
x
1
+. . . +c
n
x
n
a
i1
x
1
+. . . +a
in
x
n
≤ b
i
,
x
j
≥ 0 ,
1 ≤ i ≤ m
1 ≤ j ≤ n .
(4.10)
Maximize
subject to
λ
1
b
1
+. . . +λ
m
b
m
λ
1
a
1j
+. . . +λ
m
a
mj
≥ c
j
,
λ
i
≥ 0 ,
1 ≤ j ≤ n
1 ≤ i ≤ m .
(4.11)
Deﬁnition: Let Ω
p
= ¦x ∈ R
n
[Ax ≤ b, x ≥ 0¦ be the set of all points satisfying the constraints
of the primal problem. Similarly let Ω
d
= ¦λ ∈ R
m

A ≥ c

, λ ≥ 0¦. A point x ∈ Ω
p
(λ ∈ Ω
d
) is
said to be a feasible solution or feasible decision for the primal (dual).
The next result is trivial.
Lemma 1: (Weak duality) Let x ∈ Ω
p
, λ ∈ Ω
d
. Then
c

x ≤ λ

Ax ≤ λ

b. (4.12)
Proof: x ≥ 0 and λ

A − c

≥ 0 implies (λ

A−c

)x ≥ 0 giving the ﬁrst inequality. b−Ax ≥ 0 and
λ

≥ 0 implies λ

(b−Ax) ≥ 0 giving the second inequality. ♦
Corollary 1: If x

∈ Ω and λ

∈ Ω
d
such that c

x

= (λ

)

b, then x

is optimal for (4.10) and λ

is
optimal for (4.11).
Theorem 1: (Strong duality) Suppose Ω
p
= φ and Ω
d
= φ. Then there exists x

which is optimum
for (4.10) and λ

which is optimum for (4.11). Furthermore, c

x

= (λ

)

b.
Proof: Because of the Corollary 1 it is enough to prove the last statement, i.e., we must show that
there exist x ≥ 0, λ ≥ 0, such that Ax ≤ b, A

λ ≥ c and b

λ−c

x ≤ 0. By introducing slack
variables y ∈ R
m
, µ ∈ R
m
, r ∈ R, this is equivalent to the existence of x ≥ 0, y ≥ 0, λ ≥ 0, µ ≤
0, r ≤ 0 such that

A I
m
A

−I
n
−c

b

1
¸
¸

x
y
λ
µ
r
¸
¸
¸
¸
¸
¸
=

b
c
0
¸
¸
By the algebraic version of Farkas’ Lemma, this is possible only if
A

ξ −cθ ≤ 0 , ξ ≤ 0 ,
Aw = bθ ≤ 0 , −w ≤ 0 ,
θ ≤ 0
(4.13)
implies
b

ξ +c

w ≤ 0. (4.14)
34 CHAPTER 4. LINEAR PROGRAMMING
Case (i): Suppose (w, ξ, θ) satisﬁes (4.13) and θ < 0. Then (ξ/θ) ∈ Ω
d
, (w/−θ) ∈ Ω
p
, so that by
Lemma 1 c

w/(−θ) ≤ b

ξ/θ, which is equivalent to (4.14) since θ < 0.
Case (ii): Suppose (w, ξ, θ) satisﬁes (4.13) and θ = 0, so that −A

ξ ≥ 0, −ξ ≥ 0, Aw ≤ 0, w ≥ 0.
By hypothesis, there exist x ∈ Ω
p
, λ ∈ Ω
d
. Hence, −b

ξ = b

(−ξ) ≥ (Ax)

(−ξ) = x

(−A

ξ) ≥ 0,
and c

w ≤ (A

λ)

w = λ

(Aw) ≤ 0. So that b

ξ +c

w ≤ 0. ♦
The existence part of the above result can be strengthened.
Theorem 2: (i) Suppose Ω
p
= φ. Then there exists an optimum decision for the primal LP iff

d
= φ.
(ii) Suppose Ω
d
= φ. Then there exists an optimum decision for the dual LP iff Ω
p
= φ.
Proof Because of the symmetry of the primal and dual it is enough to prove only (i). The
sufﬁciency part of (i) follows from Theorem 1, so that only the necessity remains. Suppose, in
d
= φ. We will show that sup ¦c

x[x ∈ Ω
p
¦ = +∞. Now, Ω
d
= φ means
there does not exist λ ≥ 0 such that A

λ ≥ c. Equivalently, there does not exist λ ≥ 0, µ ≤ 0 such
that
¸
A

[
[
−I
n

λ
−−−
µ
¸
¸
=

c

By Farkas’ Lemma there exists w ∈ R
n
such that Aw ≤ 0, −w ≤ 0, and c

w > 0. By hypothesis,

p
= φ, so there exists x ≥ 0 such that Ax ≤ b. but then for any θ > 0, A(x + θw) ≤ b,
(x + θw) ≥ 0, so that (x + θw) ∈ Ω
p
. Also, c

(x + θw) = c

x + θc

w. Evidently then, sup
¦c

x[x ∈ Ω
p
¦ = +∞so that there is no optimal decision for the primal. ♦
Remark: In Theorem 2(i), the hypothesis that Ω
p
= φ is essential. Consider the following exercise.
Exercise 3: Exhibit a pair of primal and dual problems such that neither has a feasible solution.
Theorem 3: (Optimality condition) x

∈ Ω
p
is optimal if and only if there exists λ

∈ Ω
d
such that
m
¸
j=1
a
ij
x

j
< b
i
implies λ

i
= 0 ,
and
m
¸
i=1
λ

i
a
ij
< c
j
implies x

j
= 0 .
(4.15)
((4.15) is known as the condition of complementary slackness.)
Proof: First of all we note that for x

∈ Ω
p
, λ

∈ Ω
d
, (4.15) is equivalent to (4.16):

)

(Ax

−b) = 0, and (A

λ

−c)

x

= 0 . (4.16)
Necessity. Suppose x

∈ Ω
p
is optimal. Then from Theorem 2, Ω
d
= φ, so that by Theorem 1
there exists λ

∈ Ω
d
such that c

x

= (λ

)

b. By Lemma 1 we always have
c

x

≤ (λ

)

Ax

≤ (λ

)

b so that we must have c

x

= (λ

)

Ax

= (λ

)

b. But (4.16) is just an
equivalent rearrangement of these two equalities.
Sufﬁciency. Suppose (4.16) holds for some x

∈ Ω
p
, λ

∈ Ω
d
. The ﬁrst equality in (4.16) yields

)

b = (λ

)

Ax

= (A

λ

)

x

, while the second yields (A

λ

)

x

= c

x

, so that c

x

= (λ

)

b.
By Corollary 1, x

is optimal. ♦
4.2. QUALITATIVE THEORY OF LINEAR PROGRAMMING 35
The conditions x

∈ Ω
p
, x

∈ Ω
d
in Theorem 3 can be replaced by the weaker x

≥ 0, λ

≥ 0
provided we strengthen (4.15) as in the following result, whose proof is left as an exercise.

≥ 0 is optimal for the primal if and only if there exists λ

≥ 0 such
that
L(x, λ

) ≤ L(x

, λ

) ≤ L(x

, λ) for all x ≥ 0, and allλ ≥ 0, (4.17)
where L: R
n
xR
m
→ R is deﬁned by
L(x, λ) = c

x −λ

(Ax −b) (4.18)
Exercise 4: Prove Theorem 4.
Remark. The function L is called the Lagrangian. A pair (x

, λ

) satisfying (4.17) is said to form
a saddle-point of L over the set ¦x[x ∈ R
n
, x ≥ 0¦ ¦λ[λ ∈ R
m
, λ ≥ 0¦.
4.2.2 Results for problem (4.9).
It is possible to derive analogous results for LPs of the form (4.9). We state these results as exercises,
indicating how to use the results already obtained. We begin with a pair of LPs:
Maximize
subject to
c
1
x
1
+. . . +c
n
x
n
a
il
x
1
+. . . +a
in
x
n
= b
i
,
x
j
≥ 0 ,
1 ≤ i ≤ m ,
1 ≤ j ≤ n .
(4.19)
Minimize
subject to
λ
1
b
1
+. . . +λ
m
b
m
λ
1
a
1j
+. . . +λ
m
a
mj
≥ c
j
, 1 ≤ j ≤ n .
(4.20)
Note that in (4.20) the λ
i
are unrestricted in sign. Again (4.19) is called the primal and (4.20) the
dual. We let Ω
p
, Ω
d
denote the set of all x, λ satisfying the constraints of (4.19), (4.20) respectively.
Exercise 5: Prove Theorems 1 and 2 with Ω
p
and Ω
d
interpreted as above. (Hint. Replace (4.19)
by the equivalent LP: maximize c

x, subject to Ax ≤ b, (−A)x ≤ (−b), x ≥ 0. This is now of the
form (4.10). Apply Theorems 1 and 2.)
Exercise 6: Show that x

∈ Ω
p
is optimal iff there exists λ

∈ Ω
d
such that
x

j
> 0 implies
m
¸
i=1
λ

i
a
ij
= c
j
.
Exercise 7: x

≥ 0 is optimal iff there exists λ

∈ R
m
such that
L(x, λ

) ≤ L(x

, λ

) ≤ L(x

, λ) for all x ≥ 0, λ ∈ R
m
.
where L is deﬁned in (4.18). (Note that, unlike (4.17), λ is not restricted in sign.)
Exercise 8: Formulate a dual for (4.7), and obtain the result analogous to Exercise 5.
36 CHAPTER 4. LINEAR PROGRAMMING
4.2.3 Sensitivity analysis.
We investigate how the maximum value of (4.10) or (4.19) changes as the vectors b and c change.
The matrix Awill remain ﬁxed. Let Ω
p
and Ω
d
be the sets of feasible solutions for the pair (4.10) and
(4.11) or for the pair (4.19) and (4.20). We write Ω
p
(b) and Ω
d
(c) to denote the explicit dependence
on b and c respectively. Let B = ¦b ∈ R
m
[Ω
p
(b) = φ¦ and C = ¦c ∈ R
n
[Ω
d
(c) = φ¦, and for
(b, c) ∈ B C deﬁne
M(b, c) = max ¦c

x[x ∈ Ω
p
(b)¦ = min ¦λ

b[λ ∈ Ω
d
(c)¦ . (4.21)
For 1 ≤ i ≤ m, ε ∈ R, b ∈ R
m
denote
b(i, ε) = (b
1
, b
2
, . . . , b
i−1
, b
i
+ε, b
i+1
, . . . , b
m
)

,
and for 1 ≤ j ≤ n, ε ∈ R, c ∈ R
n
denote
c(j, ε) = (c
1
, c
2
, . . . , c
j−1
, c
j
+ε, c
j+1
, . . . , c
n
)

.
We deﬁne in the usual way the right and left hand partial derivatives of M at a point (
ˆ
b, ˆ c) ∈ BC
as follows:
∂M
+
∂b
i
(
ˆ
b, ˆ c) = lim
ε → 0
ε > 0
1
ε
¦M(
ˆ
b(i, ε), ˆ c) −M(
ˆ
b, ˆ c)¦ ,
∂M

∂b
i
(
ˆ
b, ˆ c) = lim
ε → 0
ε > 0
1
ε
¦M(
ˆ
b, ˆ c) −M(
ˆ
b(i, −ε), ˆ c)¦ ,
∂M
+
∂c
j
(
ˆ
b, ˆ c) = lim
ε → 0
ε > 0
1
ε
¦M(
ˆ
b, ˆ c(j, ε)) −M(
ˆ
b, ˆ c¦ ,
∂M

∂c
j
(
ˆ
b, ˆ c) = lim
ε → 0
ε > 0
1
ε
¦M(
ˆ
b, ˆ c −M(
ˆ
b, ˆ c(j, −ε))¦ ,
Let

B,

C denote the interiors of B, C respectively.
Theorem 5: At each (
ˆ
b, ˆ c) ∈

B

C, the partial derivatives given above exist. Furthermore, if
ˆ x ∈ Ω
p
(
ˆ
b),
ˆ
λ ∈ Ω
d
(ˆ c) are optimal, then
∂M
+
∂b
i
(
ˆ
b, ˆ c) ≤
ˆ
λ
i

∂M

∂b
i
(
ˆ
b, ˆ c) , 1 ≤ i ≤ m , (4.22)
4.3. THE SIMPLEX ALGORITHM 37
∂M
+
∂c
j
(
ˆ
b, ˆ c) ≥ ˆ x
j

∂M

∂c
j
(
ˆ
b, ˆ c) , 1 ≤ j ≤ n , (4.23)
Proof: We ﬁrst show (4.22), (4.23) assuming that the partial derivatives exist. By strong duality
M(
ˆ
b, ˆ c) =
ˆ
λ

ˆ
b, and by weak duality M(
ˆ
b(i, ε), ˆ c) ≤
ˆ
λ

ˆ
b(i, ε), so that
1
ε
¦M(
ˆ
b(i, ε), ˆ c) −M(
ˆ
b, ˆ c)¦ ≤
1
ε
ˆ
λ

¦
ˆ
b(i, ε) −
ˆ

ˆ
λ
i
, for ε > 0,
1
ε
¦M(
ˆ
b, ˆ c) −M(
ˆ
b(i, −ε), ˆ c)¦ ≥
1
ε
ˆ
λ

¦
ˆ
b −
ˆ
b(i, −ε)¦ =
ˆ
λ
i
, for ε > 0.
Taking limits as ε → 0, ε > 0, gives (4.22).
On the other hand, M(
ˆ
b, ˆ c) = ˆ c

ˆ x, and M(
ˆ
b, ˆ c(j, ε)) ≥ (ˆ c(j, ε))

ˆ x, so that
1
ε
¦M(
ˆ
b, ˆ c(j, ε)) −M(
ˆ
b, ˆ c)¦ ≥
1
ε
¦ˆ c(j, ε)

− ˆ c¦

ˆ x = ˆ x
j
, for ε > 0,
1
ε
¦M(
ˆ
b, ˆ c) −M(
ˆ
b, ˆ c(j, −ε))¦ ≤
1
ε
¦ˆ c − ˆ c(j, −ε)¦

ˆ x = ˆ x
j
, for ε > 0,
which give (4.23) as ε → 0, ε > 0.
Finally, the existence of the right and left partial derivatives follows from Exercises 8, 9 below. ♦
We recall some fundamental deﬁnitions from convex analysis.
Deﬁnition: X ⊂ R
n
is said to be convex if x, y ∈ X and 0 ≤ θ ≤ 1 implies (θx+(1−θ)y) ∈ X.
Deﬁnition: Let X ⊂ R
n
and f : X → R. (i) f is said to be convex if X is convex, and x, y ∈ X,
0 ≤ θ ≤ 1 implies f(θx + (1 − θ)y) ≤ θf(x) + (1 − θ)f(y). (ii) f is said to be concave if −f is
convex, i.e., x, y ∈ X, 0 ≤ θ ≤ 1 implies f(θx + (1 −θ)y) ≥ θf(x) + (1 −θ)f(y).
Exercise 8: (a) Show that Ω
p
, Ω
d
, and the sets B ⊂ R
m
, C ⊂ R
n
deﬁned above are convex sets.
(b) Show that for ﬁxed c ∈ C, M(, c) : B → R is concave and for ﬁxed b ∈ B, M(b, ) : C → R
is convex.
Exercise 9: Let X ⊂ R
n
, and f : X → R be convex. Show that at each point ˆ x in the interior of
X, the left and right hand partial derivatives of f exist. (Hint: First show that for
ε
2
> ε
1
> 0 > δ
1
> δ
2
,(1/ε
2
)¦f(ˆ x(i, ε
2
)) −f(ˆ x)¦ ≥ (1/ε
1
)¦f(ˆ x(i, ε
1
)) −f(ˆ x))¦ ≥
(1/δ
1
)¦f(ˆ x(i, δ
1
)) −f(ˆ x))¦ ≥ (1/δ
2
)¦f(ˆ x(i, δ
2
)) −f(ˆ x)¦. Then the result follows immediately.)
Remark 1: Clearly if (∂M/∂b
i
)(
ˆ
b) exists, then we have equality in (4.22), and then this result
compares with 3.14).
Remark 2: We can also show without difﬁculty that M(, c) and M(b, ) are piecewise linear (more
accurately, linear plus constant) functions on B and C respectively. This is useful in some
computational problems.
Remark 3: The variables of the dual problem are called Lagrange variables or dual variables or
shadow-prices. The reason behind the last name will be clear in Section 4.
4.3 The Simplex Algorithm
4.3.1 Preliminaries
We now present the celebrated Simplex algorithm for ﬁnding an optimum solution to any LP of the
form (4.24):
Maximize
subject to
c
1
x
1
+. . . +c
n
x
n
a
il
x
1
+. . . +a
in
x
n
= b
i
,
x
j
≥ 0 ,
1 ≤ i ≤ m
1 ≤ j ≤ n .
(4.24)
38 CHAPTER 4. LINEAR PROGRAMMING
As mentioned in 4.1 the algorithm rests upon the observations that if an optimal exists, then at least
one vertex of the feasible set Ω
p
is an optimal solution. Since Ω
p
has only ﬁnitely many vertices (see
Corollary 1 below), we only have to investigate a ﬁnite set. The practicability of this investigation
depends on the ease with which we can characterize the vertices of Ω
p
. This is done in Lemma 1.
In the following we let A
j
denote the jth column of A, i.e., A
j
= (a
1j
, . . . , a
mj
)

. We begin with
a precise deﬁnition of a vertex.
Deﬁnition: x ∈ Ω
p
is said to be a vertex of Ω
p
if x = λy + (1 −λ)z, with y, z in Ω
p
and
0 < λ < 1, implies x = y = z.
Deﬁnition: For x ∈ Ω
p
, let I(x) = ¦j[x
j
> 0¦.
Lemma 1: Let x ∈ Ω
p
. Then x is a vertex of Ω
p
iff ¦A
j
[j ∈ I(x)¦ is a linearly independent set.
Exercise 1: Prove Lemma 1.
Corollary 1: Ω
p
has at most
m
¸
j=1
n!
(n −j)!
vertices.
Lemma 2: Let x

be an optimal decision of (4.24). Then there is a vertex z

of Ω
p
which is optimal.
Proof: If ¦A
j
[j ∈ I(x

)¦ is linearly independent, let z

= x

and we are done. Hence suppose
¦A
j
[j ∈ I(x

)¦ is linearly dependent so that there exist γ
j
, not all zero, such that
¸
j∈I(x

)
γ
j
A
j
= 0 .
For θ ∈ R deﬁne z(θ) ∈ R
n
by
z
j
(θ) =

x

j
= θγ
j
,
x

j
= 0 ,
j ∈ I(x

)
j ∈ I(x

) .
Az(θ) =
¸
j∈I(x

)
z
j
(θ)A
j
=
¸
j∈I(x

)
x

j
A
j

¸
j∈I(x

)
γ
j
A
j
= b +θ 0 = b .
Since x

j
> 0 for j ∈ I(x

), it follows that z(θ) ≥ 0 when
[θ[ ≤ min

x

j

j
|

j ∈ I(x

)
¸
= θ

say .
Hence z(θ) ∈ Ω
p
whenever [θ[ ≤ θ

. Since x

is optimal we must have
c

x

≥ c

z(θ) = c

x

¸
j∈I(x

)
c
j
y
j
for −

θ ≤ θ ≤ θ

.
Since θ can take on positive and negative values, the inequality above can hold on if
¸
J∈I(x

)
c
j
γ
j
=
0, and then c

x

= c

z(θ), so that z(θ) is also an optimal solution for [θ[ ≤ θ

. But from the
deﬁnition of z(θ) it is easy to see that we can pick θ
0
with [θ
0
[ = θ

such that z
j

0
) = x

j

0
γ
j
= 0
for at least one j = j
0
in I(x

). Then,
I(z(θ
0
)) ⊂ I(x

) −¦j
0
¦ .
4.3. THE SIMPLEX ALGORITHM 39
Again, if ¦A
j
[j ∈ I(z(θ
0
))¦ is linearly independent, then we let z

= z(θ
0
) and we are done.
Otherwise we repeat the procedure above with z(θ
0
). Clearly, in a ﬁnite number of steps we will
ﬁnd an optimal decision z

which is also vertex. ♦
At this point we abandon the geometric term “vertex” and how to established LP terminology.
Deﬁnition: (i) z is said to be a basic feasible solution if z ∈ Ω
p
, and ¦A
j
[j ∈ I(z)¦ is linearly
independent. The set I(z) is then called the basis at z, and x
j
, j ∈ I(z), are called the basic
variables at z. x
j
, j ∈ I(z) are called the non-basic variables at z.
Deﬁnition: A basic feasible solution z is said to be non-degenerate if I(z) has m elements.
Notation: Let z be a non-degenerate basic feasible solution, and let j
1
< j
2
< . . . < j
m
constitute I(z). Let D(z) denote the m m non-singular matrix D(z) = [A
j
1
.
.
.A
j
2
.
.
. . . .
.
.
.A
jm
], let
c(z) denote the m-dimensional column vector c(z) = (c
j
1
, . . . , c
jm
)

and deﬁne λ(z) by λ

(z) =
c

(z)[D(z)]
−1
. We call λ(z) the shadow-price vector at z.
Lemma 3: Let z be a non-degenerate basic feasible solution. Then z is optimal if and only if
λ

(z)A ≥ c
j
, for all , j ∈ I(z) . (4.25)
Proof: By Exercise 6 of Section 2.2, z is optimal iff there exists λ such that
λ

A
j
= c
j
, for , j ∈ I(z) , (4.26)
λ

A
j
≥ c
j
, for , j ∈ I(z) , (4.27)
But since z is non-degenerate, (4.26) holds iff λ = λ(z) and then (4.27) is the same as (4.25). ♦
4.3.2 The Simplex Algorithm.
The algorithm is divided into two parts: In Phase I we determine if Ω
p
is empty or not, and if not,
we obtain a basic feasible solution. Phase II starts with a basic feasible solution and determines if
it is optimal or not, and if not obtains another basic feasible solution with a higher value. Iterating
on this procedure, in a ﬁnite number of steps, either we obtain an optimum solution or we discover
that no optimum exists, i.e., sup ¦c

x[x ∈ Ω
p
¦ = +∞. We shall discuss Phase II ﬁrst.
We make the following simplifying assumption. We will comment on it later.
Assumption of non-degeneracy. Every basic feasible solution is non-degenerate.
Phase II:
Step 1. Let z
0
be a basic feasible solution obtained from Phase I or by any other means. Set k = 0
and go to Step 2.
Step 2. Calculate [D(z
k
)]
−1
,c(z
k
), and the shadow-price vector λ

(z
k
) = c

(z
k
)[D(z
k
)]
−1
. For
each j ∈ I(z
k
) calculate c
j
−λ

(z
k
)A
j
. If all these numbers are ≤ 0, stop, because z
k
is optimal
by Lemma 3. Otherwise pick any
ˆ
j ∈ I(z
k
) such that c
ˆ
j
−λ

(z
k
)A
ˆ
j
> 0 and go to Step 3.
Step 3. Let I(z
k
) consist of j
1
< j
2
< . . . < j
m
. Compute the vector
γ
k
= (γ
k
j
1
, . . . γ
k
jm
)

= [D(z
k
)]
−1
A
ˆ
j
. If γ
k
≤ 0, stop, because by Lemma 4 below, there is no
ﬁnite optimum. Otherwise go to Step 4.
Step 4. Compute θ = min ¦(z
k
j
γ
k
j
)[j ∈ i(z), γ
k
j
> 0¦. Evidently 0 < θ < ∞. Deﬁne z
k+1
by
40 CHAPTER 4. LINEAR PROGRAMMING
z
k+1
j
=

z
k
j
−θγ
k
j
θ
z
k
j
= 0
,
,
,
j ∈ I(z)
j =
ˆ
j
j =
ˆ
j and j ∈ I(z) .
(4.28)
By Lemma 5 below, z
k+1
is a basic feasible solution with c

z
k+1
> c

z
k
. Set k = k + 1 and return
to Step 2.
Lemma 4: If γ
k
≤ 0, sup ¦c

x[x ∈ Ω
p
¦ = ∞.
Proof: Deﬁne z(θ) by
z
j
(θ) =

z
j
−θγ
k
j
θ
z
j
= 0
,
,
,
j ∈ I(z)
j =
ˆ
j
j ∈ I(z) and j =
ˆ
j .
(4.29)
First of all, since γ
k
≤ 0 it follows that z(θ) ≥ 0 for θ ≥ 0. Next, Az(θ) = Az − θ
¸
j∈I(z)
γ
k
j
A
j
+
θA
ˆ
j
= Az by deﬁnition of γ
k
. Hence, z(θ) ∈ Ω
p
for θ ≥ 0. Finally,
c

z(θ)
= c

z −θc

(z
k

k
+θc
ˆ
j
= c

z +θ¦c
ˆ
j
−c

(z
k
)[D(z
k
)]
−1
A
ˆ
j
¦
= c

z +θ¦c
ˆ
j
−λ

(z
k
)A
ˆ
j
¦i .
(4.30)
But from step 2 ¦c
ˆ
j −λ

(z
k
)A
ˆ
j
¦ > 0, so that c

z(θ) → ∞as θ → ∞. ♦
Lemma 5: z
k+1
is a basic feasible solution and c

z
k+1
> c

z
k
.
Proof: Let
˜
j ∈ I(z
k
) be such that γ
k
˜
j
> 0 and z
k
˜
j
= θγ
k
˜
j
. Then from (4.28) we see that z
k+1
˜
j
= 0,
hence
I(z
k+1
) ⊂ (I(z) −¦
˜
j¦)
¸
¦
ˆ
j¦ , (4.31)
so that it is enough to prove that A
˜
j
is independent of ¦A
j
[j ∈ I(z), j =
˜
j¦. But if this is not the
case, we must have γ
k
˜
j
= 0, giving a contradiction. Finally if we compare (4.28) and (4.29), we see
from (4.30) that
c

z
k+1
−c

z
k
= θ¦c
ˆ
j
−γ

(z
k
)A
ˆ
j
¦ ,
which is positive from Step 2. ♦
Corollary 2: In a ﬁnite number of steps Phase II will obtain an optimal solution or will determine
that sup¦c

x[x ∈ Ω
p
¦ = ∞.
Corollary 3: Suppose Phase II terminates at an optimal basic feasible solution z

. Then γ(z

) is an
optimal solution of the dual of (4.24).
Exercise 2: Prove Corollaries 2 and 3.
Remark 1: By the non-degeneracy assumption, I(z
k+1
) has m elements, so that in (4.31) we must
have equality. We see then that D(z
k+1
) is obtained from D(z
k
) by replacing the column A
j
by
4.3. THE SIMPLEX ALGORITHM 41
the column A
ˆ
j
. More precisely if D(z
k
) = [A
j
1
.
.
. . . .
.
.
.A
j
i−1
.
.
.A
˜
j
.
.
.A
j
i+1
.
.
. . . .
.
.
.A
jm
] and if
j
k
<
ˆ
j < j
k+1
then D(z
k+1
) = [A
j
1
.
.
. . . .
.
.
.A
j
i−1
.
.
.A
j
i+1
.
.
. . . .
.
.
.A
j
k
.
.
.A
ˆ
j
.
.
.A
j
k+1
.
.
. . . .
.
.
.A
jm
]. Let E be the
matrix E = [A
j
1
.
.
. . . .
.
.
.A
j
i−1
.
.
.A
ˆ
j
.
.
.A
j
i+1
.
.
. . . .
.
.
.A
jm
]. Then [D(z
k+1
)]
−1
= P E
−1
where the matrix P
permutes the columns of D(z
k+1
) such that E = D(z
k+1
)P. Next, if A
ˆ
j
=
m
¸
=1
γ
j

A
j

, it is easy
to check that E
−1
= M[D(z
k
)]
−1
where
M =

1
1
.
.
.
1
−γ
j
1
γ
˜
j
1
γ
˜
j
−γ
jm
γ
˜
j
1
.
.
.
1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

ith column
Then [D(z
k+1
)]
−1
= PM[D(z
k
)]
−1
, so that these inverses can be easily computed.
Remark 2: The similarity between Step 2 of Phase II and Step 2 of the algorithm in 3.3.4 is
striking. The basic variables at z
k
correspond to the variables w
k
and non-basic variables
correspond to u
k
. For each j ∈ I(z
k
) we can interpret the number c
j
−λ

(z
k
)A
j
to be the net
increase in the objective value per unit increase in the jth component of z
k
. This net increase is due
to the direct increase c
j
minus the indirect decrease λ

(z
k
)A
j
due to the compensating changes in
the basic variables necessary to maintain feasibility. The analogous quantity in 3.3.4 is
(∂f
0
/∂u
j
)(x
k
) −(λ
k
)

(∂f/∂u
j
)(x
k
).
Remark 3: By eliminating any dependent equations in (4.24) we can guarantee that the matrix A
has rank n. Hence at any degenerate basic feasible solution z
k
we can always ﬁnd
¯
I(z
k
) ⊃ I(z
k
)
such that
¯
I(z
k
) has m elements and ¦A
j
[j ∈
¯
I(z
k
)¦ is a linearly independent set. We can apply
Phase II using
¯
I(z
k
k
). But then in Step 4 it may turn out that θ = 0 so that
z
k+1
= z
k
. The reason for this is that
¯
I(z
k
) is not unique, so that we have to try various
alternatives for
¯
I(z
k
) until we ﬁnd one for which θ > 0. In this way the non-degeneracy
assumption can be eliminated. For details see (Canon, et al., [1970]).
We now describe how to obtain an initial basic feasible solution.
Phase I:
Step I. by multiplying some of the equality constraints in (4.24) by −1 if necessary, we can assume
that b ≥ 0. Replace the LP (4.24) by the LP (4.32) involving the variables x and y:
Maximize −
m
¸
i=1
y
i
subject to a
il
x
1
+. . . +a
in
x
n
+y
i
= b
i
, 1 ≤ i ≤ m ,
x
j
≥ 0 , y
i
≥ 0 , 1 ≤ j ≤ n , 1 ≤ i ≤ m .
(4.32)
42 CHAPTER 4. LINEAR PROGRAMMING
Go to step 2.
Step 2. Note that (x
0
, y
0
) = (0, b) is a basic feasible solution of (4.32). Apply phase II to (4.32)
starting with this solution. Phase II must terminate in an optimum based feasible solution (x

, y

)
since the value of the objective function in (4.32) lies between −
m
¸
i=1
b
i
and 0. Go to Step 3.
Step 3. If y

= 0, x

is a basic feasible solution for (4.24). If y

= 0, by Exercise 3 below, (4.24)
has no feasible solution.
Exercise 3: Show that (4.24) has a feasible solution iff y

= 0.
4.4 LP Theory of a Firm in a Competitive Economy
4.4.1 Activity analysis of the ﬁrm.
We think of a ﬁrm as a system which transforms input into outputs. There are m kinds of inputs
and k kinds of outputs. Inputs are usually classiﬁed into raw materials such as iron ore, crude oil,
or raw cotton; intermediate products such as steel, chemicals, or textiles; capital goods
3
such as
machines of various kinds, or factory buildings, ofﬁce equipment, or computers; ﬁnally various
kinds of labor services. The ﬁrm’s outputs themselves may be raw materials (if it is a mining
company) or intermediate products (if it is a steel mill) or capital goods (if it manufactures lathes)
or ﬁnished goods (if it makes shirts or bakes cookies) which go directly to the consumer. Labor is
not usually considered an output since slavery is not practiced; however, it may be considered an
output in a “closed,” dynamic Malthusian framework where the increase in labor is a function of the
output. (See the von Neumann model in (Nikaido [1968]), p. 141.)
Within the ﬁrm, this transformation can be conducted in different ways, i.e., different combina-
tions of inputs can be used to produce the same combination of outputs, since human labor can
do the same job as some machines and machines can replace other kinds of machines, etc. This
substitutability among inputs is a fundamental concept in economics. We formalize it by specifying
which transformation possibilities are available to the ﬁrm.
By an input vector we mean any m-dimensional vector r = (r
1
, . . . , r
m
)

with r ≥ 0, and by an
output vector we mean any k-dimensional vector y = (y
1
, . . . , y
k
)

with y ≥ 0. We now make three
(i) The transformation of inputs into outputs is organized into a ﬁnite number, say n, of processes
or activities.
(ii) Each activity combines the k inputs in ﬁxed proportions into the m outputs in ﬁxed propor-
tions. Furthermore, each activity can be conducted at any non-negative intensity or level. Pre-
cisely, the jth activity is characterized completely by two vectors A
j
= (a
1j
, a
2j
, . . . , a
mj
)

and
B
j
= (b
ij
, . . . , b
kj
)

so that if it is conducted at a level x
j
≥ 0, then it combines (transforms) the
input vector (a
1j
x
j
, . . . , a
mj
x
j
)

= x
j
A
j
into the output vector (b
1j
x
j
, . . . , b
kj
x
j
)

= x
j
B
j
. Let
A be the mn matrix [A
1
.
.
. . . .
.
.
.A
n
] and B be the k n matrix B = [B
1
.
.
. . . .
.
.
.B
n
].
3
It is more accurate to think of the services of capital goods rather than these goods themselves as inputs. It is these
services which are consumed in the transformation into outputs.
4.4. LP THEORY OF A FIRM IN A COMPETITIVE ECONOMY 43
(iii) If the ﬁrmconducts all the activities simultaneously with the jth activity at level x
j
≥ 0, 1 ≤ j ≤
n, then it transforms the input vector x
1
A
1
+. . . +x
n
A
n
into the output vector x
1
B
1
+. . . +x
n
B
n
.
With these assumptions we know all the transformations technically possible as soon as we spec-
ify the matrices A and B. Which of these possible transformations will actually take place depends
upon their relative proﬁtability and availability of inputs. We study this next.
4.4.2 Short-term behavior.
In the short-term, the ﬁrm cannot change the amount available to it of some of the inputs such as
capital equipment, certain kinds of labor, and perhaps some raw materials. Let us suppose that these
inputs are 1, 2, . . . , and they are available in the amounts r

1
, . . . , r

, whereas the supply of the
remaining inputs can be varied. We assume that the ﬁrm is operating in a competitive economy
which means that the unit prices p = (p
1
, . . . , p
k
)

of the outputs, and q = (q
1
, . . . , q
m
)

of the
inputs is ﬁxed. Then the manager of the ﬁrm, if he is maximizing the ﬁrm’s proﬁts, faces the
following decision problem:
Maximize p

y −
m
¸
j=+1
q
j
r
j
subject to y = Bx,
a
i1
x
1
+. . . +a
in
x
n
≤ r

i
, 1 ≤ i ≤ ,
a
i1
x
1
+. . . +a
in
x
n
≤ r
i
, + 1 ≤ i ≤ m ,
x
j
≥ 0, 1 ≤ j ≤ n; r
i
≥ 0 , + 1 ≤ i ≤ m .
(4.33)
The decision variables are the activity levels x
1
, . . . , x
n
, and the short-term input supplies r
+1
, . . . , r
m
.
The coefﬁcients of B and Aare the ﬁxed technical coefﬁcients of the ﬁrm, the r

i
are the ﬁxed short-
term supplies, whereas the p
i
, q
j
are prices determined by the whole economy, which the ﬁrm ac-
cepts as given. Under realistic conditions (4.33) has an optimal solution, say, x

1
, . . . , x

n
, r

+1
, . . . , r

m
.
4.4.3 Long-term equilibrium behavior.
In the long run the supplies of the ﬁrst inputs are also variable and the ﬁrm can change these
supplies from r

1
, . . . , r

by buying or selling these inputs at the market price q
1
, . . . , q

. Whether
the ﬁrm will actually change these inputs will depend upon whether it is proﬁtable to do so, and in
turn this depends upon the prices p, q. We say that the prices (p

, q

) and a set of input supplies
r

= (r

1
, . . . , r

m
) are in (long-term) equilibrium if the ﬁrm has no proﬁt incentive to change r

under the prices (p

, q

).
Theorem 1: p

, q

, r

are in equilibrium if and only if q

is an optimal solution of (4.34):
Minimize (r

)

q
subject to A

q ≥ B

p

q ≥ 0 .
(4.34)
Proof: Let c = B

p

. By deﬁnition, p

, q

, r

are in equilibrium iff for all ﬁxed ∆ ∈ R
m
,
M(∆) ≤ M(0) where M(∆) is the maximum value of the LP (4.35):
Maximize c

x −(q

)

subject to Ax ≤ r

+ ∆ ,
x ≥ 0 .
(4.35)
44 CHAPTER 4. LINEAR PROGRAMMING
For ∆ = 0, (4.34) becomes the dual of (4.35) so that by the strong duality theorem, M(0) = (r

)

q

.
Hence p

, q

, r

are in equilibrium iff
c

x −(q

)

∆ ≤ M(0) = (r

)

q

, (4.36)
whenever x is feasible for (4.35). By weak duality if x is feasible for (4.35) and q is feasible for
(4.34),
c

x −(q

)

∆ ≤ q

(r

= ∆) −(q

)

∆ , (4.37)
and, in particular, for q = q

,
c

x −(q

)

∆ ≤ (q

)

(r

+ ∆) −(q

)

∆ = (q

)

r

Remark 1: We have shown that (p

, q

, r

are in long-term equilibrium iff q

is an optimum
solution to the dual (namely (4.34)) of (4.38):
Maximize c

x
subject to Ax ≤ r

x ≥ 0 .
(4.38)
This relation between p

, q

, r

has a very nice economic interpretation. Recall that c = B

p

, i.e.,
c
j
= p

1
b
1j
+p

2
b
2j
+. . . +p

k
b
kj
. Now b
ij
is the amount of the ith output produced by operating the
jth activity at a unit level x
j
= 1. Hence, c
j
is the revenue per unit level operation of the jth activity
so that c

x is the revenue when the n activities are operated at levels x. On the other hand if the jth
activity is operated at level x
j
= 1, it uses an amount a
ij
of the ith input. If the ith input is valued at
a

i
, then the input cost of operating at x
j
= 1, is
m
¸
i=1
q
i
a
ij
, so that the input cost of operating the n
activities at levels x is (A

q

)

= (q

)

Ax. Thus, if x

is the optimum activity levels for (4.38) then
the output revenue is c

x

and the input cost is (q

)

Ax

. But from (4.16), (q

)

(Ax

−r

) = 0 so
that
c

x

= (q

)

r

, (4.39)
i.e., at the optimum activity levels, in equilibrium, total revenues = total cost of input supplies. In
fact, we can say even more. From (4.15) we see that if x
=
ast
j
> 0 then
c
j
=
m
¸
i=1
q

i
a
ij
,
i.e., at the optimum, the revenue of an activity operated at a positive level = input cost of that activity.
Also if
c
j
<
m
¸
i=1
q

i
a
ij
,
then x

j
= 0, i.e., if the revenue of an activity is less than its input cost, then at the optimum it is
operated at zero level. Finally, again from (4.15), if an equilibrium the optimum ith input supply r

i
is greater than the optimum demand for the ith input,
4.4. LP THEORY OF A FIRM IN A COMPETITIVE ECONOMY 45
r

i
>
n
¸
j=1
a
ij
x

j
,
then q

i
= 0, i.e., the equilibrium price of an input which is in excess supply must be zero, in other
words it must be a free good.
Remark 2: Returning to the short-term decision problem (4.33), suppose that

1
, . . . , λ

, λ

+1
, . . . , λ

m
) is an optimum solution of the dual of (4.33). Suppose that the market
prices of inputs 1, . . . , are q
1
, . . . , q

. Let us denote by M(∆
1
, . . . , ∆

) the optimum value of
(4.33) when the amounts of the inputs in ﬁxed supply are r

1
+ ∆
1
, . . . , r

+ ∆

. Then if
(∂M/∂∆
i
)[
∆=0
exists, we can see from (4.22) that it is always proﬁtable to increase the ith input
i
if λ

i
> q
i
, and conversely it is proﬁtable to sell some
of the ith input at price q
i
if λ

i
< q
i
. Thus λ

i
can be interpreted as the ﬁrm’s internal valuation of
the ith input or the ﬁrm’s imputed or shadow price of the ith input. This interpretation has wide
applicability, which we mention brieﬂy. Often engineering design problems can be formulated as
LPs of the form (4.10) or (4.19), where some of the coefﬁcients b
i
are design parameters. The
design procedure is to ﬁx these parameters at some nominal value b

i
, and carry out the
optimization problem. Suppose the resulting optimal dual variables are λ

i
. then we see (assuming
differentiability) that it is worth increasing b

i
if the unit cost of increasing this parameter is less
than λ

i
, and it is worth decreasing this parameter if the reduction in total cost per unit decrease is
greater than λ

i
.
4.4.4 Long-term equilibrium of a competitive, capitalist economy.
The proﬁt-maximizing behavior of the ﬁrm presented above is one of the two fundamental building
blocks in the equilibrium theory of a competitive, capitalist economy. Unfortunately we cannot
present the details here. We shall limit ourselves to a rough sketch. We think of the economy as
a feedback process involving ﬁrms and consumers. Let us suppose that there are a total of h com-
modities in the economy including raw materials, intermediate and capital goods, labor, and ﬁnished
products. By adding zero rows to the matrices (A, B) characterizing a ﬁrm we can suppose that all
the h commodities are possible inputs and all the h commodities are possible outputs. Of course,
for an individual ﬁrm most of the inputs and most of the outputs will be zero. the sole purpose for
making this change is that we no longer need to distinguish between prices of inputs and prices of
outputs. We observe the economy starting at time T. At this time there exists within the economy
an inventory of the various commodities which we can represent by a vector ω = (ω
1
, . . . , ω
h
) ≥ 0.
ω is that portion of the outputs produced prior to T which have not been consumed up to T. We are
assuming that this is a capitalist economy, which means that the ownership of ω is divided among
the various consumers j = 1, . . . , J. More precisely, the jth consumer owns the vector of commodi-
ties ω(j) ≥ 0, and
J
¸
j=1
ω(j) = ω. We are including in ω(j) the amount of his labor services which
consumer j is willing to sell. Now suppose that at time T the prevailing prices of the h commodities
are λ = (λ
1
, . . . , λ
h
)

≥ 0. Next, suppose that the managers of the various ﬁrms assume that the
prices λ are not going to change for a long period of time. Then, from our previous analysis we
know that the manager of the ith ﬁrm will plan to buy input supplies r(i) ≥ 0, r(i) ∈ R
h
, such
46 CHAPTER 4. LINEAR PROGRAMMING
that (λ, r(i)) is in long term equilibrium, and he will plan to produce an optimum amount, say y(i).
Here i = 1, 2, . . . , I, where I is the total number of ﬁrms. We know that r(i) and y(i) depend on
λ, so that we explicitly write r(i, λ), y(i, λ). We also recall that (see (4.38))
λ

r(i, λ) = λ

y(i, λ) , 1 ≤ i ≤ I . (4.40)
Now the ith manager can buy r(i) from only two sources: outputs from other ﬁrms, and the con-
sumers who collectively own ω. Similarly, the ith manager can sell his planned output y(i) either as
input supplies to other ﬁrms or to the consumers. Thus, the net supply offered for sale to consumers
is S(λ), where
S(λ) =
J
¸
j=1
ω(j) +
I
¸
i=1
y(i, λ) −
i
¸
i=1
r(i, λ) . (4.41)
We note two important facts. First of all, from (4.40), (4.41) we immediately conclude that
λ

S(λ) =
J
¸
j=1
λ

ω(j) , (4.42)
that is the value of the supply offered to consumers is equal to the value of the commodities (and
labor) which they own. The second point is that there is no reason to expect that S(λ) ≥ 0.
Now we come to the second building block of equilibrium theory. The value of the jth consumer’s
possessions is λ

ω(j). The theory assumes that he will plan to buy a set of commodities d(j) =
(d
1
(j), . . . , d
h
(j)) ≥ 0 so as to maximize his satisfaction subject to the constraint λ

d(j) = λ

ω(j).
Here also d(j) will depend on λ, so we write d(j, λ). If we add up the buying plans of all the
consumers we obtain the total demand
D(λ) =
J
¸
j=1
d(j, λ) ≥ 0 , (4.43)
which also satisﬁes
λ

D(λ) =
J
¸
j=1
λ

ω(j) . (4.44)
The most basic question of equilibrium theory is to determine conditions under which there exists a
price vector λ
E
such that the economy is in equilibrium, i.e., S(λ
E
) = D(λ
E
), because if such an
equilibrium price λ
E
exists, then at that price the production plans of all the ﬁrms and the buying
plan of all the consumers can be realized. Unfortunately we must stop at this point since we cannot
proceed further without introducing some more convex analysis and the ﬁxed point theorem. For
a simple treatment the reader is referred to (Dorfman, Samuelson, and Solow [1958], Chapter 13).
For a much more general mathematical treatment see (Nikaido [1968], Chapter V).
4.5.1 Some mathematical tricks.
It is often the case in practical decision problems that the objective is not well-deﬁned. There may
be a number of plausible objective functions. In our LP framework this situation can be formulated
as follows. The constraints are given as usual by Ax ≤ b, x ≥ 0. However, there are, say, k
objective functions (c
1
)

x, . . . , (c
k
)

x. It is reasonable then to deﬁne a single objective function
f
0
(x) by f
0
(x) = minimum ¦(c
1
)

x, (c
2
)

x, . . . , (c
k
)

x¦, so that we have the decision problem,
Maximize f
0
(x)
subject to Ax ≤ b, x ≥ 0 .
(4.45)
This is not a LP since f
0
is not linear. However, the following exercise shows how to transform
(4.45) into an equivalent LP.
Exercise 1: Show that (4.45) is equivalent to (4.46) below, in the sense that x

is optimal for (4.45)
iff (x

, y

) = (x

, f
0
(x

)) is optimal for (4.46).
Maximize y
subject to Ax ≤ b, x ≤ 0
y ≤ (c
i
)

x , 1 ≤ i ≤ k .
(4.46)
Exercise 1 will also indicate how to do Exercise 2.
Exercise 2: Obtain an equivalent LP for (4.47):
Maximize
n
¸
j=1
c
i
(x
i
)
subject to Ax ≤ b, x ≤ 0 ,
(4.47)
where c
i
: R → R are concave, piecewise-linear functions of the kind shown in Figure 4.3.
The above-given assumption of the concavity of the c
i
is crucial. In the next exercise, the inter-
pretation of “equivalent” is purposely left ambiguous.
Exercise 3: Construct an example of the kind (4.47), where the c
i
are piecewise linear (but not
concave), and such that there is no equivalent LP.
It turns out however, that even if the c
i
are not concave, an elementary modiﬁcation of the Simplex
algorithm can be given to obtain a “local” optimal decision. See (Miller [1963]).
4.5.2 Scope of linear programming.
LP is today the single most important optimization technique. This is because many decision prob-
lems can be adequately formulated as LPs, and, given the capabilities of modern computers, the
Simplex method (together with its variants) is an extremely powerful technique for solving LPs in-
volving thousands of variables. To obtain a feeling for the scope of LP we refer the reader to the
book by one of the originators of LP (Dantzig [1963]).
48 CHAPTER 4. LINEAR PROGRAMMING
.
.
.
c
i
(x
i
)
x
i
Figure 4.3: A function of the form used in Exercise 2.
Chapter 5
OPTIMIZATION OVER SETS
DEFINED BY INEQUALITY
CONSTRAINTS: NONLINEAR
PROGRAMMING
In many decision-making situations the assumption of linearity of the constraint inequalities in LP
is quite restrictive. The linearity of the objective function is not restrictive as shown in the ﬁrst
exercise below. In Section 1 we present the general nonlinear programming problem (NP) and
prove the Kuhn-Tucker theorem. Section 2 deals with Duality theory for the case where appropriate
convexity conditions are satisﬁed. Two applications are given. Section 3 is devoted to the important
special case of quadratic programming. The last section is devoted to computational considerations.
5.1 Qualitative Theory of Nonlinear Programming
5.1.1 The problem and elementary results.
The general NP is a decision problem of the form:
Maximize f
0
(x)
subject to (x) ≤ 0 , i = 1, . . . , m,
(5.1)
where x ∈ R
n
, f
i
: R
n
→ R, i = 0, 1, . . . , m, are differentiable functions. As in Chapter 4,
x ∈ R
n
is said to be a feasible solution if it satisﬁes the constraints of (5.1), and Ω ⊂ R
n
is the
subset of all feasible solutions; x

∈ Ω is said to be an optimal decision or optimal solution if
f
0
(x

) ≥ f
0
(x) for x ∈ Ω. From the discussion in 4.1.2 it is clear that equality constraints and sign
constraints on some of the components of x can all be transformed into the form (5.1). The next
exercise shows that we could restrict ourselves to objective functions which are linear; however, we
will not do this.
Exercise 1: Show that (5.2), with variables y ∈ R, x ∈ R
n
, is equivalent to (5.1):
49
50 CHAPTER 5. NONLINEAR PROGRAMMING
Maximize y
subject to f
i
(x) ≤ 0, 1 ≤ i ≤ m, and y −f
0
(x) ≤ 0 .
(5.2)
Returning to problem (5.1), we are interested in obtaining conditions which any optimal decision
must satisfy. The argument parallels very closely that developed in Exercise 1 of 4.1 and Exercise 1
of 4.2. The basic idea is to linearize the functions f
i
in a neighborhood of an optimal decision x

.
Deﬁnition: Let x be a feasible solution, and let I(x) ⊂ ¦1, . . . , m¦ be such that f
i
(x) = 0 for
ı ∈ I(x), f
i
(x) < 0 for i ∈ I(x). (The set I(x) is called the set of active constraints at x.)
Deﬁnition: (i) Let x ∈ Ω. A vector h ∈ R
n
is said to be an admissible direction for Ω at x if there
exists a sequence x
k
, k = 1, 2, . . . , in Ω and a sequence of numbers ε
k
, k = 1, . . . , with ε
k
> 0
for all k such that
lim
k→∞
x
k
= x ,
lim
k→∞
1
ε
k
(x
k
−x) = h .
(ii) Let C(Ω, x) = ¦h[h is an admissible direction for Ω at x¦. C(Ω, x) is called the tangent cone
of Ω at x. Let K(Ω, x) = ¦x + h[h ∈ C(Ω, x)¦. (See Figures 5.1 and 5.2 and compare them with
Figures 4.1 and 4.2.)
If we take x
k
= x and ε
k
= 1 for all k, we see that 0 ∈ C(Ω, x) so that the tangent cone is always
nonempty. Two more properties are stated below.
Exercise 2: (i) Show that C(Ω, x) is a cone, i.e., if h ∈ C(Ω, x) and θ ≥ 0, then θh ∈ C(Ω, x).
(ii) Show that C(Ω, x) is a closed subset of R
n
. (Hint for (ii): For m = 1, 2, . . . , let h
m
and
¦x
mk
, ε
mk
> 0¦

k=1
be such that x
mk
→ x and (1/ε
mk
)(x
mk
−x) → h
m
as k → ∞. Suppose
that h
m
→ h as m → ∞. Show that there exist subsequences ¦x
mkm
, ε
mkm
¦

m=1
such that
x
mkm
→ x and (1/ε
mkm
)(x
mkm
−x) → h as m → ∞.)
In the deﬁnition of C(Ω, x) we made no use of the particular functional description of Ω. The
following elementary result is more interesting in this light and should be compared with (2.18) in
Chapter 2 and Exercise 1 of 4.1.
Lemma 1: Suppose x

∈ Ω is an optimum decision for (5.1).
Then
f
0x
(x

)h ≤ 0 for all h ∈ C(Ω, x

) . (5.3)
Proof: Let x
k
∈ Ω, ε
k
> 0, k = 1, 2, 3, . . . , be such that
5.1. QUALITATIVE THEORY OF NONLINEAR PROGRAMMING 51
,
P
¦x[f
3
(x) = 0¦
Q
x

direction of
increasing
payoff
π(k) =
¦x[f
0
(x) = k¦
¦x[f
1
(x) = 0¦

R
¦x[f
2
(x) = 0¦
Figure 5.1: Ω = PQR
lim
k→∞
x
k
= x

,
lim
k→∞
1
ε
k
(x
k
−x

) = h . (5.4)
Note that in particular (5.4) implies
lim
k→∞
1
ε
k
[x
k
−x

[ = [h[ . (5.5)
Since f
0
is differentiable, by Taylor’s theorem we have
f
0
(x
k
) = f
0
(x

+ (x
k
−x

)) = f
0
(x

) +f
0x
(x

)(x
k
−x

) +o([x
k
−x

[) . (5.6)
Since x
k
∈ Ω, and x

is optimal, we have f
0
(x
k
) ≤ f
0
(x

), so that
0 ≥ f
0x
(x

)
(x
k
−x

)
ε
k
+
o(|x
k
−x

|)
ε
k
.
Taking limits as k → ∞, using (5.4) and (5.5), we can see that
0 ≥
=
lim
k→∞
f
0x
(x

)h. ♦
f
0x
(x

)
(x
k
−x

)
ε
k
+
lim
k→∞
o(|x
k
−x

|)
|x
k
−x

| lim
k→∞
|x
k
−x

|
ε
k
52 CHAPTER 5. NONLINEAR PROGRAMMING
-
-
-
-
-
-
-
-
x

K(Ω, x

)
0
C(Ω, x

)
Figure 5.2: C(Ω, x

) is the tangent cone of Ω at x

.
The basic problem that remains is to characterize the set C(Ω, x

) in terms of the derivatives of the
functions f
i
. Then we can apply Farkas’ Lemma just as in Exercise 1 of 4.2.
Lemma 2: Let x

∈ Ω. Then
C(Ω, x

) ⊂ ¦h[f
ix
(x

)h ≤ 0 for all i ∈ I(x

)¦ . (5.7)
Proof: Let h ∈ R
n
and x
k
∈ Ω, ε
k
> 0, k = 1, 2, . . . , satisfy (5.4). Since f
i
is differentiable, by
Taylor’s theorem we have
f
i
(x
k
) = f
i
(x

) +f
ix
(x

)(x
k
−x

) +o([x
k
−x

[) .
Since x
k
∈ Ω, f
i
(x
k
) ≤ 0, and if i ∈ I(x

), f
i
(x

) = 0, so that f
i
(x
k
) ≤ f
i
(x

). Following the
proof of Lemma 1 we can conclude that 0 ≥ f
ix
(x

)h. ♦
Lemma 2 gives us a partial characterization of C(Ω, x

). Unfortunately, in general the inclusion
sign in (5.7) cannot be reversed. The main reason for this is that the set ¦f
ix
(x

)[i ∈ I(x

)¦ is not
in general linearly independent.
Exercise 3: Let x ∈ R
2
, f
1
(x
1
, x
2
) = (x
1
−1)
3
+x
2
, and f
2
(x
1
, x
2
) = −x
2
. Let
(x

1
, x

2
) = (1, 0). Then I(x

) = ¦1, 2¦. Show that
5.1. QUALITATIVE THEORY OF NONLINEAR PROGRAMMING 53
C(Ω, x

) = ¦h[f
ix
(x

)h ≤ 0 , i = 1, 2, ¦.
(Note that ¦f
1x
(x

), f
2x
(x

)¦ is not a linearly independent set; see Lemma 4 below.)
5.1.2 Kuhn-Tucker Theorem.
Deﬁnition: Let x

∈ Ω. We say that the constraint qualiﬁcation (CQ) is satisﬁed at x

if
C(Ω, x) = ¦h[f
ix
(x

)h ≤ 0 for all i ∈ I(x

)¦,
and we say that CQ is satisﬁed if CQ is satisﬁed at all x ∈ Ω. (Note that by Lemma 2 C(Ω, x) is
always a subset of the right-hand side.)
Compare the next result with Exercise 2 of 4.2.
Theorem 1: (Kuhn and Tucker [1951]) Let x

be an optimum solution of (5.1), and suppose that
CQ is satisﬁed at x

. Then there exist λ

i
≥ 0, for i ∈ I(x

), such that
f
0x
(x

) =
¸
i∈I(x

)
λ

i
f
ix
(x

)
(5.8)
Proof: By Lemma 1 and the deﬁnition of CQ it follows that f
0x
(x

)h ≤ 0 whenever f
ix
(x

)h ≤ 0
for all i ∈ I(x

). By the Farkas’ Lemma of 4.2.1 it follows that there exist λ

i
≥ 0 for i ∈ I(x

)
such that (5.8) holds. ♦
In the original formulation of the decision problem we often have equality constraints of the form
r
j
(x) = 0, which get replaced by r
j
(x) ≤ 0, −r
j
(x) ≤ 0 to give the form (5.1). It is convenient in
application to separate the equality constraints from the rest. Theorem 1 can then be expressed as
Theorem 2.
Theorem 2: Consider the problem (5.9).
Maximize f
0
(x)
subject to f
i
(x) ≤ 0 , i = 1, . . . , m,
r
j
(x) = 0 , j = 1, . . . , k .
(5.9)
Let x

be an optimum decision and suppose that CQ is satisﬁed at x

. Then there exist λ

i
≥ 0, i =
1, . . . , m, and µ

j
, j = 1, . . . , k such that
f
0x
(x

) =
m
¸
i=1
λ

i
f
ix
(x

) +
k
¸
j=1
µ

j
r
jx
(x

) , (5.10)
and
λ

i
= 0 whenever f
i
(x

) < 0 . (5.11)
Exercise 4: Prove Theorem 2.
54 CHAPTER 5. NONLINEAR PROGRAMMING
An alternative form of Theorem 1 will prove useful for computational purposes (see Section 4).
Theorem 3: Consider (5.9), and suppose that CQ is satisﬁed at an optimal solution x

. Deﬁne
ψ : R
n
→ R by
ψ(h) = max ¦−f
0x
(x

)h, f
1
(x

) +f
1x
(x

)h, . . . , f
m
(x

) +f
mx
(x

)h¦ ,
and consider the decision problem
Minimize ψ(h)
subject to −ψ(h) −f
0x
(x

)h ≤ 0,
−ψ(h) +f
i
(x

) +f
ix
(x

)h ≤ 0 , 1 ≤ i ≤ m
−1 ≤ h
i
≤ 1 , i = 1, . . . , n .
(5.12)
Then h = 0 is an optimal solution of (5.12).
Exercise 5: Prove Theorem 3. (Note that by Exercise 1 of 4.5, (5.12) can be transformed into a
LP.)
Remark: For problem (5.9) deﬁne the Lagrangian function L:
(x
1
, . . . , x
n
; λ
1
, . . . , λ
m
; µ
1
, . . . , µ
k
) → f
0
(x) −
m
¸
i=1
λ
i
f
i
(x) −
k
¸
j=1
µ
j
r
j
(x).
Then Theorem 2 is equivalent to the following statement: if CQ is satisﬁed and x

is optimal, then
there exist λ

≥ 0 and µ

such that L
x
(x

, λ

, µ

) = 0 and L(x

, λ

, µ

) ≤ L(x

, λ, µ) for all
λ ≥ 0, µ.
There is a very important special case when the necessary conditions of Theorem 1 are also
sufﬁcient. But ﬁrst we need some elementary properties of convex functions which are stated as an
exercise. Some additional properties which we will use later are also collected here.
Recall the deﬁnition of convex and concave functions in 4.2.3.
Exercise 6: Let X ⊂ R
n
be convex. Let h : X → R be a differentiable function. Then
(i) h is convex iff h(y) ≥ h(x) +h
x
(x)(y −x) for all x, y, in X,
(ii) h is concave iff h(y) ≤ h(x) +h
x
(x)(y −x) for all x, y in X,
(iii) h is concave and convex iff h is afﬁne, i.e. h(x) ≡ α +b

x for some
ﬁxed α ∈ R, b ∈ R
n
.
Suppose that h is twice differentiable. Then
(iv) h is convex iff h
xx
(x) is positive semideﬁnite for all x in X,
(v) h is concave iff h
xx
(x) is negative semideﬁnite for all x in X,
(vi) h is convex and concave iff h
xx
(x) ≡ 0.
Theorem 4: (Sufﬁcient condition) In (5.1) suppose that f
0
is concave and f
i
is convex for
i = 1, . . . , m. Then
(i) Ω is a convex subset of R
n
, and
(ii) if there exist x

∈ Ω, λ

i
≥ 0, i ∈ I(x

), satisfying (5.8), then x

is an optimal solution of
(5.1).
Proof:
(i) Let y, z be in Ω so that f
i
(y) ≤ 0, f
i
(z) ≤ 0 for i = 1, . . . , m. Let 0 ≤ θ ≤ 1. Since f
i
is
convex we have
5.1. QUALITATIVE THEORY OF NONLINEAR PROGRAMMING 55
f
i
(θy + (1 −θ)z) ≤ θf
i
(y) + (1 −θ)f
i
(z) ≤ 0 , 1 ≤ i ≤ m,
so that (θy + (1 −θ)z) ∈ Ω, hence Ω is convex.
(ii) Let x ∈ Ω be arbitrary. Since f
0
is concave, by Exercise 6 we have
f
0
(x) ≤ f
0
(x

) +f
0x
(x

)(x −x

) ,
so that by (5.8)
f
0
(x) ≤ f
0
(x

) +
¸
i∈I(x

)
λ

i
f
ix
(x

)(x −x

) .
(5.13)
Next, f
i
is convex so that again by Exercise 6,
f
i
(x) ≥ f
i
(x

) +f
ix
(x

)(x −x

) ;
but f
i
(x) ≤ 0, and f
i
(x

) = 0 for i ∈ I(x

), so that
f
ix
(x

)(x −x

) ≤ 0 for i ∈ I(x

) . (5.14)
Combining (5.14) with the fact that λ

i
≥ 0, we conclude from (5.13) that f
0
(x) ≤ f
0
(x

), so that
x

is optimal. ♦
Exercise 7: Under the hypothesis of Theorem 4, show that the subset Ω

of Ω, consisting of all the
optimal solutions of (5.1), is a convex set.
Exercise 8: A function h : X → R deﬁned on a convex set X ⊂ R
n
is said to be strictly convex if
h(θy + (1 −θ)z) < θh(y) + (1 −θ)h(z) whenever 0 < θ < 1 and y, z are in X with y = z. h is
said to be strictly concave if −h is strictly convex. Under the hypothesis of Theorem 4, show that
an optimal solution to (5.1) is unique (if it exists) if either f
0
is strictly concave or if the
f
i
, 1 ≤ i ≤ m, are strictly convex. (Hint: Show that in (5.13) we have strict inequality if x = x

.)
5.1.3 Sufﬁcient conditions for CQ.
As stated, it is usually impractical to verify if CQ is satisﬁed for a particular problem. In this
subsection we give two conditions which guarantee CQ. These conditions can often be veriﬁed in
practice. Recall that a function g : R
n
→ R is said to be afﬁne if g(x) ≡ α + b

x for some ﬁxed
α ∈ R and b ∈ R
n
.
We adopt the formulation (5.1) so that
Ω = ¦x ∈ R
n
[f
i
(x) ≤ 0 , 1 ≤ i ≤ m¦ .
Lemma 3: Suppose x

∈ Ω and suppose there exists h

∈ R
n
such that for each i ∈ I(x

), either
f
ix
(x

)h

< 0, or f
ix
(x

)h

= 0 and f
i
is afﬁne. Then CQ is satisﬁed at x

.
Proof: Let h ∈ R
n
be such that f
ix
(x

)h ≤ 0 for i ∈ I(x

). Let δ > 0. We will ﬁrst show that
(h +δh

) ∈ C(Ω, x

). To this end let ε
k
> 0, k = 1, 2, . . . , be a sequence converging to 0 and set
x
k
= x

k
(h +δh

). Clearly x
k
converges to x

, and (1/ε
k
)(x
k
−x

) converges to (h +δh

).
Also for i ∈ I(x

), if f
ix
(x

)h < 0, then
f
i
(x
k
) = f
i
(x

) +ε
k
f
ix
(x

)(h +δh

) +o(ε
k
[h +δh

[)
≤ δε
k
f
ix
(x

)h

+o(ε
k
[h +δh

[)
< 0 for sufﬁciently large k ,
whereas for i ∈ I(x

), if f
i
is afﬁne, then
56 CHAPTER 5. NONLINEAR PROGRAMMING
f
i
(x
k
) = f
i
(x

) +ε
k
f
ix
(x

)(h +δh

) ≤ 0 for all k .
Finally, for i ∈ I(x

) we have f
i
(x

) < 0, so that f
i
(x
k
) < 0 for sufﬁciently large k. Thus we
have also shown that x
k
∈ Ω for sufﬁciently large k, and so by deﬁnition (h + δh

) ∈ C(Ω, x

).
Since δ > 0 can be arbitrarily small, and since C(Ω, x

) is a closed set by Exercise 2, it follows that
h ∈ C(Ω, x

). ♦
Exercise 9: Suppose x

∈ Ω and suppose there exists ˆ x ∈ R
n
such that for each i ∈ I(x

), either
f
i
(x

) < 0 and f
i
is convex, or f
i
(ˆ x) ≤ 0 and f
i
is afﬁne. Then CQ is satisﬁed at x

. (Hint: Show
that h

= ˆ x −x

satisﬁes the hypothesis of Lemma 3.)
Lemma 4: Suppose x

∈ Ω and suppose there exists h

∈ R
n
such that f
ix
(x

)h

≤ 0 for
i ∈ I(x

), and ¦f
ix
(x

)[i ∈ I(x

), f
ix
(x

)h

= 0¦ is a linearly independent set. Then CQ is
satisﬁed at x

.
Proof: Let h ∈ R
n
be such that f
ix
(x

)h ≤ 0 for all i ∈ I(x

). Let δ > 0. We will show that
(h +δh

) ∈ C(Ω, x

). Let J
δ
= ¦i[i ∈ I(x

), f
ix
(x

)(h +δh

) = 0¦, consist of p elements.
Clearly J
δ
⊂ J = ¦i[i ∈ I(x

), f
i
x(x

)h

= 0¦, so that ¦f
ix
(x

, u

)[i ∈ J
δ
¦ is linearly
independent. By the Implicit Function Theorem, there exist ρ > 0, an open set V ⊂ R
n
containing
x

= (w

, u

), and a differentiable function g : U → R
p
, where U = ¦u ∈ R
n−p
[[u −u

[ < ρ¦,
such that
f
i
(w, u) = 0, i ∈ J
δ
, and (w, u) ∈ V
iff
u ∈ U, and w = g(u) .
Next we partition h, h

as h = (ξ, η), h

= (ξ

, η

) corresponding to the partition of x = (w, u).
Let ε
k
> 0, k = 1, 2 . . . , be any sequence converging to 0, and set u
k
= u

+ ε
k
(η +δη

), w
k
=
g(u
k
), and ﬁnally x
k
= (s
k
, u
k
).
We note that u
k
converges to u

, so w
k
= g(u
k
) converges to w

= g(u

). Thus, x
k
converges
to x

. Now (1/ε
k
)(x
k
−x

) = (1/ε
k
)(w
k
−w

, u
k
−u

) = (1/ε
k
)(g(u
k
) −g(u

), ε
k
(η +δη

)).
Since g is differentiable, it follows that (1/ε
k
)(x
k
−x

) converges to (g
u
(u

)(η +δη

), η +δη

).
But for i ∈ J
δ
we have
0 = f
ix
(x

)(h +δh

) = f
iw
(x

)(ξ +δξ

) +f
iu
(x

)(η +δη

) . (5.15)
Also, for i ∈ J
δ
, 0 = f
i
(g(u), u) for u ∈ U so that 0 = f
iw
(x

)g
u
(u

) +f
iu
(x

), and hence
0 = f
iw
(x

)g
u
(u

)(η +δη

) +f
iu
(x

)(η +δη

) . (5.16)
If we compare (5.15) and (5.16) and recall that ¦f
iw
(x

)[i ∈ J
δ
¦ is a basis in R
p
we can conclude
that (ξ +δξ

) = g
u
(u

)(η +δη

) so that (1/ε
k
)(x
k
−x

) converges to (h +hδh

).
It remains to show that x
k
∈ Ω for sufﬁciently large k. First of all, for i ∈ J
δ
, f
i
(x
k
) =
f
i
(g(u
k
), u
k
) = 0, whereas for i ∈ J
δ
, i ∈ I(x

),
f
i
(x
k
) = f
i
(x

) +f
ix
(x

)(x
k
−x

) +o([x
k
−x

[)
f
i
(x

) +ε
k
f
ix
(x

)(h +δh

) +o(ε
k
) +o([x
k
−x

[),
5.2. DUALITY THEORY 57
and since f
i
(x

) = 0 whereas f
ix
(x

)(h + δh

) < 0, we can conclude that f
i
(x
k
) < 0 for sufﬁ-
ciently large k. Thus, x
k
∈ Ω for sufﬁciently large k. Hence, (h +δh

) ∈ C(Ω, x

).
To ﬁnish the proof we note that δ > 0 can be made arbitrarily small, and C(Ω, x

) is closed by
Exercise 2, so that h ∈ C(Ω, x

). ♦
The next lemma applies to the formulation (5.9). Its proof is left as an exercise since it is very
similar to the proof of Lemma 4.
Lemma 5: Suppose x

is feasible for (5.9) and suppose there exists h

∈ R
n
such that the set
¦f
ix
(x

)[i ∈ I(x

), f
ix
(x

)h

= 0¦
¸
¦r
jx
(x

)[j = 1, . . . , k¦ is linearly independent, and f
ix
(x

)h

0 for i ∈ I(x

), r
jx
(x

)h

= 0 for 1 ≤ j ≤ k. Then CQ is satisﬁed at x

.
Exercise 10: Prove Lemma 5
5.2 Duality Theory
Duality theory is perhaps the most beautiful part of nonlinear programming. It has resulted in many
applications within nonlinear programming, in terms of suggesting important computational algo-
rithms, and it has provided many unifying conceptual insights into economics and management
science. We can only present some of the basic results here, and even so some of the proofs are
relegated to the Appendix at the end of this Chapter since they depend on advanced material. How-
ever, we will give some geometric insight. In 2.3 we give some application of duality theory and in
2.2 we refer to some of the important generalizations. The results in 2.1 should be compared with
Theorems 1 and 4 of 4.2.1 and the results in 4.2.3.
It may be useful to note in the following discussion that most of the results do not require differ-
entiability of the various functions.
5.2.1 Basic results.
Consider problem (5.17) which we call the primal problem:
Maximize f
0
(x)
subject to f
i
(x) ≤
ˆ
b
i
, 1 ≤ i ≤ m
x ∈ X ,
(5.17)
where x ∈ R
n
, f
i
: R
n
→ R, 1 ≤ i ≤ m, are given convex functions, f
0
: R
n
→ R is a
given concave function, X is a given convex subset of R
n
and
ˆ
b = (
ˆ
b
1
, . . . ,
ˆ
b
m
)

is a given vector.
For convenience, let f = (f
1
, . . . , f
m
)

: R
n
→ R
m
. We wish to examine the behavior of the
maximum value of (5.17) as
ˆ
b varies. So we deﬁne
Ω(b) = ¦x[x ∈ X, f(x) ≤ b¦, B = ¦b[Ω(b) = φ¦,
and
M : B → R
¸
¦+∞¦ by M(b) = sup¦f
0
(x)[x ∈ X, f(x) ≤ b¦
= sup¦f
0
(x)[x ∈ Ω(b)¦ ,
so that in particular if x

is an optimal solution of (5.17) then M(
ˆ
b) = f
0
(ˆ x). We need to consider
the following problem also. Let λ ∈ R
m
, λ ≥ 0, be ﬁxed.
Maximize f
0
(x) −λ

(f(x) −
ˆ
b)
subject to x ∈ X ,
(5.18)
58 CHAPTER 5. NONLINEAR PROGRAMMING
and deﬁne
m(λ) = sub¦f
0
(x) −λ

(f(x) −
ˆ
b)[x ∈ X¦ .
Problem (5.19) is called the dual problem:
Minimize m(λ)
subject to λ ≥ 0 .
(5.19)
Let m

= inf ¦m(λ)[λ ≥ 0¦.
Remark 1: The set X in (5.17) is usually equal to R
n
and then, of course, there is no reason to
separate it out. However, it is sometimes possible to include some of the constraints in X in such
a way that the calculation of m(λ) by (5.18) and the solution of the dual problem (5.19) become
simple. For example see the problems discussed in Sections 2.3.1 and 2.3.2 below.
Remark 2: It is sometimes useful to know that Lemmas 1 and 2 below hold without any convexity
conditions on f
0
, f, X. Lemma 1 shows that the cost function of the dual problem is convex which
is useful information since there are computation techniques which apply to convex cost functions
but not to arbitrary nonlinear cost functions. Lemma 2 shows that the optimum value of the dual
problem is always an upper bound for the optimum value of the primal.
Lemma 1: m : R
n
+
→ R
¸
¦+∞¦ is a convex function. (Here R
n
+
= ¦λ ∈ R
n
[λ ≥ 0¦.)
Exercise 1: Prove Lemma 1.
Lemma 2: (Weak duality) If x is feasible for (5.17), i.e., x ∈ Ω(
ˆ
b), and if λ ≥ 0, then
f
0
(x) ≤ M(
ˆ
b) ≤ m

≤ m(λ) . (5.20)
Proof: Since f(x) −
ˆ
b ≤ 0, and λ ≥ 0, we have λ

(f(x) −
ˆ
b) ≤ 0. So,
f
0
(x) ≤ f
0
(x) −λ

(f(x) −
ˆ
b), for x ∈ Ω(
ˆ
b), λ ≥ 0 .
Hence
f
0
(x) ≤ sup ¦f
0
(x)[x ∈ Ω(
ˆ
b)¦ = M(
ˆ
b)
≤ sup ¦f
0
(x) −λ

(f(x) −
ˆ
b)[x ∈ Ω(
ˆ
b)¦ and since Ω(
ˆ
b) ⊂ X,
≤ sup ¦f
0
(x) −λ

(f(x) −
ˆ
b)[x ∈ X¦ = m(λ) .
Thus, we have
f
0
(x) ≤ M(
ˆ
b) ≤ m(λ) for x ∈ Ω(
ˆ
b), λ ≥ 0 ,
and since M(
ˆ
b) is independent of λ, if we take the inﬁmum with respect to λ ≥ 0 in the right-hand
inequality we get (5.20). ♦
The basic problem of Duality Theory is to determine conditions under which M(
ˆ
b) = m

in
(5.20). We ﬁrst give a simple sufﬁciency condition.
Deﬁnition: A pair (ˆ x,
ˆ
λ) with ˆ x ∈ X, and
ˆ
λ ≤ 0 is said to satisfy the optimality conditions if
5.2. DUALITY THEORY 59
ˆ x is optimal solution of (5.18) with λ =
ˆ
λ, (5.21)
ˆ x is feasible for (5.17), i.e., f
i
(ˆ x) ≤
ˆ
b
i
for i = 1, . . . , m , (5.22)
ˆ
λ
i
= 0 when f
i
(ˆ x) <
ˆ
b
i
, equivalently,
ˆ
λ

(f(ˆ x) −
ˆ
b) = 0. (5.23)
ˆ
λ ≥ 0 is said to be an optimal price vector if there is ˆ x ∈ X such that (ˆ x,
ˆ
λ) satisfy the optimality
condition. Note that in this case ˆ x ∈ Ω(
ˆ
b) by virtue of (5.22).
The next result is equivalent to Theorem 4(ii) of Section 1 if X = R
n
, and f
i
, 0 ≤ i ≤ m, are
differentiable.
Theorem 1: (Sufﬁciency) If (ˆ x,
ˆ
λ) satisfy the optimality conditions, then ˆ x is an optimal solution to
the primal,
ˆ
λ is an optimal solution to the dual, and M(
ˆ
b) = m

.
Proof: Let x ∈ Ω(
ˆ
b), so that
ˆ
λ

(f(x) −
ˆ
b) ≤ 0. Then
f
0
(x) ≤ f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b)
≤ sup¦f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b)[x ∈ X¦
= f
0
(ˆ x) −
ˆ
λ

(f(ˆ x) −
ˆ
b) by (5.21)
= f
0
(ˆ x) by (5.23)
so that ˆ x is optimal for the primal, and hence by deﬁnition f
0
(ˆ x) = M(
ˆ
b). Also
m(
ˆ
λ) = f
0
(ˆ x) −
ˆ
λ

(f(ˆ x) −
ˆ
b)
f
0
(ˆ x) = M(
ˆ
b) ,
so that from Weak Duality
ˆ
λ is optimal for the dual. ♦
We now proceed to a much more detailed investigation.
Lemma 3: B is a convex subset of R
m
, and M : B → R
¸
¦+∞¦ is a concave function.
Proof: Let b,
˜
b belong to B, let x ∈ Ω(b), ˜ x ∈ Ω(
˜
b), let 0 ≤ θ ≤ 1. Then (θx + (1 − θ)˜ x) ∈ X
since X is convex, and
f
i
(θx + (1 −θ)˜ x) ≤ θf
i
(x) + (1 −θ)f
i
(˜ x)
since f
i
is convex, so that
f
i
(θx + (1 −θ)˜ x) ≤ θb + (1 −θ)
˜
b , (5.24)
hence
(θx + (1 −θ)˜ x) ∈ Ω(θb + (1 −θ)
˜
b)
and therefore, B is convex.
Also, since f
0
is concave,
f
0
(θx + (1 −θ)˜ x) ≥ θf
0
(x) + (1 −θ)f
0
(˜ x) .
60 CHAPTER 5. NONLINEAR PROGRAMMING
Since (5.24) holds for all x ∈ Ω(b) and ˜ x ∈ Ω(
˜
b) it follows that
M(θb + (1 −θ)
ˆ
b) ≥ sup ¦f
0
(θx + (1 −θ)˜ x)[x ∈ Ω(b), ˜ x ∈ Ω(
˜
b)¦
≥ sup¦f
0
(x)[x ∈ Ω(b)¦ + (1 −θ) sup ¦f
0
(˜ x)[˜ x ∈ Ω(
˜
b)¦
= θM(b) + (1 −θ)M(
˜
b). ♦
Deﬁnition: Let X ⊂ R
n
and let g : X → R
¸
¦∞, −∞¦. A vector λ ∈ R
n
is said to be a
g(x) ≤ g(ˆ x) +λ

(x − ˆ x) for x ∈ X.
(g(x) ≥ g(ˆ x) +λ

(x − ˆ x) for x ∈ X.)
(See Figure 5-3.)
,
-
-
-
-
-
.
.
.
.
M(b)
M(
ˆ
b)
b ∈ B
ˆ
b b
M is not stable at
ˆ
b
M(b)
M(
ˆ
b)
ˆ
b
b
M is stable at
ˆ
b
M(b)
M(
ˆ
b) +λ

(b −
ˆ
b)
ˆ
b b
ˆ
b
Figure 5.3: Illustration of supergradient of stability.
Deﬁnition: The function M : B → R
¸
¦∞¦ is said to be stable at
ˆ
b ∈ B if there exists a real
number K such that
M(b) ≤ M(
ˆ
b) +K[b −
ˆ
b[ for b ∈ B .
(In words, M is stable at
ˆ
b if M does not increase inﬁnitely steeply in a neighborhood of
ˆ
b. See
Figure 5.3.)
A more geometric way of thinking about subgradients is the following. Deﬁne the subset A ⊂
R
1+m
by
5.2. DUALITY THEORY 61
A = ¦(r, b)[b ∈ B, and r ≤ M(b)¦ .
Thus A is the set lying ”below” the graph of M. We call A the hypograph
1
of M. Since M is
concave it follows immediately that A is convex (in fact these are equivalent statements).
Deﬁnition: A vector (λ
0
, λ
1
, . . . , λ
m
) is said to be the normal to a hyperplane supporting A at a
point(ˆ r,
ˆ
b) if
λ
0
ˆ r +
m
¸
i=1
λ
i
ˆ
b
i
≥ λ
0
r +
m
¸
i=1
λ
i
b
i
for all (r, b) ∈ A . (5.25)
(In words, Alies below the hyperplane ˆ π = ¦(r, b)[λ
0
r+
¸
λ
i
b
i
= λ
0
ˆ r+
¸
λ
i
b
i
¦.) The supporting
hyperplane is said to be non-vertical if λ
0
= 0. See Figure 5.4.
Exercise 2: Show that if
ˆ
b ∈ B,
˜
b ≥
ˆ
b, and ˜ r ≤ M(
ˆ
b), then
˜
b ∈ B, M(
˜
b), and (˜ r,
˜
b) ∈ A.
Exercise 3: Assume that
ˆ
b ∈ B, and M(
ˆ
b) < ∞. Show that (i) if λ = (λ
1
, . . . , λ
m
)

is a
ˆ
b then λ ≥ 0, and (1, −λ
1
, . . . , −λ
m
)

deﬁnes a non-vertical hyperplane
supporting A at (M(
ˆ
b),
ˆ
b), (ii) if (λ
0
, −λ
1
, . . . , −λ
m
)

deﬁnes a hyperplane supporting A at
(M(
ˆ
b),
ˆ
b) then λ
0
≥ 0, λ
i
≥ 0 for 1 ≤ i ≤ m; futhermore, if the hyperplane is non-vertical then
((λ
1

0
, . . . , (λ
m

0
))

is a supergradient of M at
ˆ
b.
We will prove only one part of the next crucial result. The reader who is familiar with the
Separation Theorem of convex sets should be able to construct a proof for the second part based
on Figure 5.4, or see the Appendix at the end of this Chapter.
Lemma 4: (Gale [1967]) M is stable at
ˆ
b iff M has a supergradient at
ˆ
b. Proof: (Sufﬁciency only)
Let λ be a supergradient at
ˆ
b, then
M(b) ≤ M(
ˆ
b) +λ

(b −
ˆ
b)
≤ M(
ˆ
b) +[λ[[b −
ˆ
b[ . ♦
The next two results give important alternative interpretations of supergradients.
Lemma 5: Suppose that ˆ x is optimal for (5.17). Then
ˆ
λ is a supergradient of M at
ˆ
b iff
ˆ
λ is an
optimal price vector, and then (ˆ x,
ˆ
λ) satisfy the optimality conditions.
Proof: By hypothesis, f(ˆ x) = M(
ˆ
b), ˆ x ∈ X, and f(ˆ x) ≤
ˆ
b. Let
ˆ
λ be a supergradient of M at
ˆ
b.
By Exercise 2, (M(
ˆ
b), f(ˆ x)) ∈ A and by Exercise 3,
ˆ
λ ≥ 0 and
M(
ˆ
b) −
ˆ
λ

ˆ
b ≥ M(
ˆ
b) −
ˆ
λ

f(ˆ x) ,
so that
ˆ
λ

(f(ˆ x) −
ˆ
b) ≥ 0. But then
ˆ
λ

(
ˆ
b − f(ˆ x)) = 0, giving (5.23). Next let x ∈ X. Then
(f
0
(x), f(x)) ∈ A, hence again by Exercise 3
M(
ˆ
b) −
ˆ
λ

ˆ
b ≥ f
0
(x) −
ˆ
λ

f(x) .
Since f
0
(ˆ x) = M(
ˆ
b), and
ˆ
λ

(f(ˆ x) −
ˆ
b) = 0, we can rewrite the inequality above as
1
From the Greek “hypo” meaning below or under. This neologism contrasts with the epigraph of a function which is
the set lying above the graph of the function.
62 CHAPTER 5. NONLINEAR PROGRAMMING
.
.
M(
ˆ
b)
M(b)
ˆ
b
A
b
No non-vertical hyperplane supporting A at (M(
ˆ
b),
ˆ
b)

0
, . . . , λ
m
)
M(
ˆ
b)
M(b)
ˆ π
A
b
ˆ
b
ˆ π is a non-vertical hyperplane supporting A at (M(
ˆ
b),
ˆ
b)
Figure 5.4: Hypograph and supporting hyperplane.
f
0
(ˆ x) +
ˆ
λ

(f(ˆ x) −
ˆ
b) ≥ f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b) ,
so that (5.21) holds. It follows that (ˆ x,
ˆ
λ) satisfy the optimality conditions.
Conversely, suppose ˆ x ∈ X,
ˆ
λ ≥ 0 satisfy (5.21), (5.22), and (5.23). Let x ∈ Ω(b), i.e.,
x ∈ X, f(x) ≤ b. Then
ˆ
λ

(f(x) −b) ≤ 0 so that
f
0
(x) ≤ f
0
(x)
ˆ
λ

(f(x) −b)
= f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b) +
ˆ
λ

(b −
ˆ
b)
≤ f
0
(ˆ x) −
ˆ
λ

(f(ˆ x) −
ˆ
b) +
ˆ
λ

(b −
ˆ
b) by (5.21)
= f
0
(ˆ x) +
ˆ
λ

(b −
ˆ
b) by (5.23)
= M(
ˆ
b) +
ˆ
λ

(b −
ˆ
b) .
Hence
M(b) = sup¦f
0
(x)[x ∈ Ω(b)¦ ≤ M(
ˆ
b) +
ˆ
λ

(b −
ˆ
b) ,
so that
ˆ
λ

is a supergradient of M at
ˆ
b. ♦
Lemma 6: Suppose that
ˆ
b ∈ B, and M(
ˆ
b) < ∞. Then
ˆ
λ is a supergradient of M at
ˆ
b iff
ˆ
λ is an
optimal solution of the dual (5.19) and m(
ˆ
λ) = M(
ˆ
b).
Proof: Let
ˆ
λ be a supergradient of M at
ˆ
b. Let x ∈ X. By Exercises 2 and 3
5.2. DUALITY THEORY 63
M(
ˆ
b) −
ˆ
λ

ˆ
b ≥ f
0
(x) −
ˆ
λ

f(x)
or
M(
ˆ
b) ≥ f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b) ,
so that
M(
ˆ
b) ≥ sup¦f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b)[x ∈ X¦ = m(
ˆ
λ) .
By weak duality (Lemma 2) it follows that M(
ˆ
b) = m(
ˆ
λ) and
ˆ
λ is optimal for (5.19).
Conversely suppose
ˆ
λ ≥ 0, and m(
ˆ
λ) = M(
ˆ
b). Then for any x ∈ X
M(
ˆ
b) ≥ f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b) ,
and if moreover f(x) ≤ b, then
ˆ
λ

(f(x) −b) ≤ 0, so that
M(
ˆ
b) ≥ f
0
(x) −
ˆ
λ

(f(x) −
ˆ
b) +
ˆ
λ

(f(x) −b)
= f
0
(x) −
ˆ
λ

b +
ˆ
λ

ˆ
b for x ∈ Ω(b) .
Hence,
M(b) = sup¦f
0
(x)[x ∈ Ω(b)¦ ≤ M(
ˆ
b) +
ˆ
λ

(b −
ˆ
b) ,
so that
ˆ
We can now summarize our results as follows.
Theorem 2: (Duality) Suppose
ˆ
b ∈ B, M(
ˆ
b) < ∞, and M is stable at
ˆ
b. Then
(i) there exists an optimal solution
ˆ
λ for the dual, and m(
ˆ
λ) = M(
ˆ
b),
(ii)
ˆ
λ is optimal for the dual iff
ˆ
λ is a supergradient of M at
ˆ
b,
(iii) if
ˆ
λ is any optimal solution for the dual, then ˆ x is optimal for the primal iff (ˆ x,
ˆ
λ) satisfy the
optimality conditions of (5.21), (5.22), and (5.23).
Proof: (i) follows from Lemmas 4,6. (ii) is implied by Lemma 6. The “if” part of (iii) follows from
Theorem 1, whereas the “only if” part of (iii) follows from Lemma 5. ♦
Corollary 1: Under the hypothesis of Theorem 2, if
ˆ
λ is an optimal solution to the dual then
(∂M
+
/∂b
i
)(
ˆ
b) ≤
ˆ
λ
i
≤ (∂M

/∂b
i
)(
ˆ
b).
Exercise 4: Prove Corollary 1. (Hint: See Theorem 5 of 4.2.3.)
5.2.2 Interpretation and extensions.
It is easy to see using convexity properties that, if X = R
n
and f
i
, 0 ≤ i ≤ m, are differentiable,
then the optimality conditions (5.21), (5.22), and (5.23) are equivalent to the Kuhn-Tucker condition
(5.8). Thus the condition of stability of M at
ˆ
b plays a similar role to the constraint qualiﬁcation.
However, by Lemmas 4, 6 stability is equivalent to the existence of optimal dual variables, whereas
CQ is only a sufﬁcient condition. In other words if CQ holds at ˆ x then M is stable at
ˆ
b. In particular,
if X = R
n
and the f
i
are differentiable, the various conditions of Section 1.3 imply stability. Here
we give one sufﬁcient condition which implies stability for the general case.
Lemma 7: If
ˆ
b is in the interior of B, in particular if there exists x ∈ X such that f
i
(x) <
ˆ
b
i
for
1 ≤ i ≤ m, then M is stable at
ˆ
b.
64 CHAPTER 5. NONLINEAR PROGRAMMING
The proof rests on the Separation Theorem for convex sets, and only depends on the fact that M
is concave, M(
ˆ
b) < ∞ without loss of generality, and
ˆ
b is the interior of B. For details see the
Appendix.
Much of duality theory can be given an economic interpretation similar to that in Section 4.4.
Thus, we can think of x as the vector of n activity levels, f
0
(x) the corresponding revenue, X as
constraints due to physical or long-term limitations, b as the vector of current resource supplies,
and ﬁnally f(x) the amount of these resources used up at activity levels x. The various convexity
conditions are generalizations of the economic hypothesis of non-increasing returns-to-scale. The
primal problem (5.17) is the short-term decision problem faced by the ﬁrm. Next, if the current
resources can be bought or sold at prices
ˆ
λ = (λ
1
, . . . , λ
m
)

, the ﬁrm faces the decision problem
(5.18). If for a price system
ˆ
λ, an optimal solution of (5.17) also is an optimal solution for (5.18),
then we can interpret
ˆ
λ as a system of equilibrium prices just as in 4.2. Assuming the realistic
condition
ˆ
b ∈ B, M(
ˆ
b) < ∞ we can see from Theorem 2 and its Corollary 1 that there exists
an equilibrium price system iff (∂M
+
/∂b
i
)(
ˆ
b) < ∞, 1 ≤ i ≤ m; if we interpret (∂M
+
/∂b
i
)(
ˆ
b)
as the marginal revenue of the ith resource, we can say that equilibrium prices exist iff marginal
productivities of every (variable) resource is ﬁnite. These ideas are developed in (Gale [1967]).
.
M(b)
A
b
ˆ
b
M(
ˆ
b)
Figure 5.5: If M is not concave there may be no supporting hyperplane at (M(
ˆ
b),
ˆ
b).
Referring to Figure 5.3 or Figure 5.4, and comparing with Figure 5.5 it is evident that if M is not
concave or, equivalently, if its hypograph Ais not convex, there may be no hyperplane supporting A
at (M(
ˆ
b),
ˆ
b). This is the reason why duality theory requires the often restrictive convexity hypoth-
esis on X and f
i
. It is possible to obtain the duality theorem under conditions slightly weaker than
convexity but since these conditions are not easily veriﬁable we do not pursue this direction any fur-
ther (see Luenberger [1968]). A much more promising development has recently taken place. The
basic idea involved is to consider supporting A at (M(
ˆ
b),
ˆ
b) by (non-vertical) surfaces ˆ π more gen-
eral than hyperplanes; see Figure 5.6. Instead of (5.18) we would then have more general problem
of the form (5.26):
Maximize f
0
(x) −F(f(x) −
ˆ
b)
subject to x ∈ X ,
(5.26)
5.2. DUALITY THEORY 65
where F : R
m
→ R is chosen so that ˆ π (in Figure 5.6) is the graph of the function b → M(
ˆ
b) −
F(b −
ˆ
b). Usually F is chosen from a class of functions φ parameterized by µ = (µ
1
, . . . , µ
k
) ≥ 0.
Then for each ﬁxed µ ≥ 0 we have (5.27) instead of (5.26):
Maximize f
0
(x) −φ(µ; f(x) −
ˆ
b)
subject to x ∈ X .
(5.27)
.
M(b)
ˆ π
A
b
ˆ
b
M(
ˆ
b)
Figure 5.6: The surface ˆ π supports A at (M(
ˆ
b),
ˆ
b).
If we let
ψ(µ) =sup¦f
0
(x) −φ(µ; f(x) −
ˆ
b)[x ∈ X¦ .
then the dual problem is
Minimize ψ(µ)
subject to µ ≥ 0 ,
in analogy with (5.19).
The economic interpretation of (5.27) would be that if the prevailing (non-uniform) price system
is φ(µ; ) then the resources f(x) −
ˆ
b can be bought (or sold) for the amount φ(µ; f(x) −
ˆ
b). For
such an interpretation to make sense we should have φ(µ; b) ≥ 0 for b ≥ 0, and φ(µ; b) ≥ φ(µ;
˜
b)
whenever b ≥
˜
b. A relatively unnoticed, but quite interesting development along these lines is
presented in (Frank [1969]). Also see (Arrow and Hurwicz [1960]).
For non-economic applications, of course, no such limitation on φ is necessary. The following
references are pertinent: (Gould [1969]), (Greenberg and Pierskalla [1970]), (Banerjee [1971]). For
more details concerning the topics of 2.1 see (Geoffrion [1970a]) and for a mathematically more
elegant treatment see (Rockafellar [1970]).
5.2.3 Applications.
Decentralized resource allocation.
Parts (i) and (iii) of Theorem 2 make duality theory attractive for computation purposes. In particular
from Theorem 2 (iii), if we have an optimal dual solution
ˆ
λ then the optimal primal solutions are
those optimal solutions of (5.18) for λ =
ˆ
λ which also satisfy the feasibility condition (5.22) and
66 CHAPTER 5. NONLINEAR PROGRAMMING
the “complementary slackness” condition (5.23). This is useful because generally speaking (5.18)
is easier to solve than (5.17) since (5.18) has fewer constraints.
Consider a decision problem in a large system (e.g., a multi-divisional ﬁrm). The system is
made up of k sub-systems (divisions), and the decision variable of the ith sub-system is a vector
x
i
∈ R
n
i
, 1 ≤ i ≤ k. The sub-system has individual constraints of the form x
i
∈ X
i
where x
i
is
a convex set. Furthermore, the sub-systems share some resources in common and this limitation is
expressed as f
1
(x
1
) +. . . +f
k
(x
k
) ≤
ˆ
b where f
i
: R
n
i
→ R
m
are convex functions and
ˆ
b ∈ R
m
is the vector of available common resources. Suppose that the objective function of the large system
is additive, i.e. it is the form f
1
0
(x
1
) + . . . + f
k
0
(x
k
) where f
i
0
: R
n
i
→ R are concave functions.
Thus we have the decision problem (5.28):
Maximize
k
¸
i=1
f
i
0
(x
i
)
subject to x
i
∈ X
i
, 1 ≤ i ≤ k,
k
¸
i=1
f
i
(x
i
) ≤
ˆ
b .
(5.28)
For λ ∈ R
m
, λ ≥ 0, the problem corresponding to (5.19) is
Maximize f
i
0
(x
i
) −λ

f
i
(x
i
) −λ

(
k
¸
i=1
f
i
(x
i
) −
ˆ
b)
subject to x
i
∈ X
i
, 1 ≤ i ≤ k ,
which decomposes into k separate problems:
Maximize f
i
0
(x
i
) −λ

f
i
(x
i
)
subject to x
i
∈ X
i
, 1 ≤ i ≤ k .
(5.29)
If we let m
i
(λ) = sup¦f
i
0
(x
i
) −λ

f
i
(x
i
)[x
i
∈ X
i
¦, and m(λ) =
k
¸
i=1
m
i
(λ) +λ

ˆ
b, then the dual
problem is
Minimize m(λ) ,
subject to λ ≥ 0 .
(5.30)
Note that (5.29) may be much easier to solve than (5.28) because, ﬁrst of all, (5.29) involves fewer
constraints, but perhaps more importantly the decision problems in (5.29) are decentralized whereas
in (5.28) all the decision variables x
1
, . . . , x
k
are coupled together; in fact, if k is very large it may
be practically impossible to solve (5.28) whereas (5.29) may be trivial if the dimensions of x
i
are
small.
Assuming that (5.28) has an optimal solution and the stability condition is satisﬁed, we need to
ﬁnd an optimal dual solution so that we can use Theorem 2(iii). For simplicity suppose that the
f
i
0
, 1 ≤ i ≤ k, are strictly concave, and also suppose that (5.29) has an optimal solution for every
λ ≥ 0. Then by Exercise 8 of Section 1, for each λ ≥ 0 there is a unique optimal solution of (5.29),
say x
i
(λ). Consider the following algorithm.
5.2. DUALITY THEORY 67
Step 1. Select λ
0
≥ 0 arbitrary. Set p = 0, and go to Step 2.
Step 2. Solve (5.29) for λ = λ
p
and obtain the optimal solution x
p
= (x
1

p
), . . . , x
k

p
)).
Compute e
p
=
k
¸
i=1
f
i
(x
i

p
)) −
ˆ
b. If e
p
≥ 0, x
p
is feasible for (5.28) and can easily be seen to be
optimal.
Step 3. Set λ
p=1
according to
λ
p+1
i
=

λ
p
i
if e
p
i
≥ 0
λ
p
i
−d
p
e
p
i
if e
p
i
< 0
where d
p
> 0 is chosen a priori. Set p = p + 1 and return to Step 3.
It can be shown that if the step sizes d
p
are chosen properly, x
p
will converge to the optimum
solution of (5.28). For more detail see (Arrow and Hurwicz [1960]), and for other decentralization
schemes for solving (5.28) see (Geoffrion [1970b]).
Control of water quality in a stream.
The discussion in this section is mainly based on (Kendrick, et al., [1971]). For an informal discus-
sion of schemes of pollution control which derive their effectiveness from duality theory see (Solow
[1971]). See (Dorfman and Jacoby [1970].)
Figure 5.7 is a schematic diagram of a part of a stream into which n sources (industries and
municipalities) discharge polluting efﬂuents. The pollutants consist of various materials, but for
simplicity of exposition we assume that their impact on the quality of the stream is measured in
terms of a single quantity, namely the biochemical oxygen demand (BOD) which they place on the
dissolved oxygen (DO) in the stream. Since the DO in the stream is used to breakdown chemically
the pollutants into harmless substances, the quality of the stream improves with the amount of
DO and decreases with increasing BOD. It is a well-advertized fact that if the DO drops below a
certain concentration, then life in the stream is seriously threatened; indeed, the stream can “die.”
Therefore, it is important to treat the efﬂuents before they enter the stream in order to reduce the
BOD to concentration levels which can be safely absorbed by the DO in the stream. In this example
we are concerned with ﬁnding the optimal balance between costs of waste treatment and costs of
high BOD in the stream.
We ﬁrst derive the equations which govern the evolution in time of BOD and DO in the n areas
of the streams. The ﬂuctuations of BOD and DO will be cyclical with a period of 24 hours. Hence,
it is enough to study the problem over a 24-hour period. We divide this period into T intervals,
t = 1, . . . , T. During interval t and in area i let
z
i
(t) = concentration of BOD measured in mg/liter,
q
i
(t) = concentration of DO measured in mg/liter,
s
i
(t) = concentration of BOD of efﬂuent discharge in mg/liter, and
m
i
(t) = amount of efﬂuent discharge in liters.
The principle of conservation of mass gives us equations (5.31) and (5.32):
z
i
(t + 1) −z
i
(t) = −α
i
z
i
(t) +
ψ
i−1
z
i−1
(t)
v
i

ψ
i
z
i
(t)
v
i
+
s
i
(t)m
i
(t)
v
i
, (5.31)
q
i
(t + 1) −q
i
(t) = β
i
(q
s
i
−q
i
(t)) +
ψ
i−1
q
i−1
(t)
v
i

ψ
i
q
i
(t)
v
i

i
z
i
(t) −η
i
v
i
, t = 1, . . . , T and i = 1, . . . , N.
(5.32)
68 CHAPTER 5. NONLINEAR PROGRAMMING
. . . . . .
. . .
direction of ﬂow
0
z
0
q
0
1
z
1
q
1
i −1
z
i−1
q
i−1
i
z
i
q
i
i + 1
z
i+1
q
i+1
N
z
N
q
N
N + 1
given
(1 −π
1
)s
i
s
i
1 −π
i−1
s
i−a
s
i
(1 −π
i
)s
i
s
i+1
(1 −π
i+1
)s
i+1
s
N
(1 −π
N
)s
N
Figure 5.7: Schematic of stream with efﬂuent discharges.
Here, v
i
= volume of water in area i measured in liters, ψ
i
= volume of water which ﬂows from
area i to are i +1 in each period measured in liters. α
i
is the rate of decay of BOD per interval. This
decay occurs by combination of BOD and DO. β
i
is the rate of generation of DO. The increase in
DO is due to various natural oxygen-producing biochemical reactions in the stream and the increase
is proportional to (q
s
− q
i
) where q
s
is the saturation level of DO in the stream. Finally, η
i
is the
DO requirement in the bottom sludge. The v
i
, ψ
i
, α
i
, η
i
, q
s
are parameters of the stream and are
assumed known. They may vary with the time interval t. Also z
0
(t), q
0
(t) which are the concen-
trations immediately upstream from area 1 are assumed known. Finally, the initial concentrations
z
i
(1), q
i
(1), i = 1, . . . , N are assumed known.
Now suppose that the waste treatment facility in area i removes in interval t a fraction π
i
(t) of
the concentration s
i
(t) of BOD. Then (5.31) is replaced by
z
i
(t + 1) −z
i
(t) = −α
i
z
i
(t) +
ψ
i
z
i−1
v
i

ψ
i
z
i
(t)
v
i
+
(1−π
i
(t))s
i
(t)m
i
(t)
v
i
. (5.33)
We now turn to the costs associated with waste treatment and pollution. The cost of waste treat-
ment can be readily identiﬁed. In period t the ith facility treats m
i
(t) liters of efﬂuent with a BOD
concentration s
i
(t) mg/liter of which the facility removes a fraction π
i
(t). Hence, the cost in period
t will be f
i

i
(t), s
i
(t), m
i
(t)) where the function must be monotonically increasing in all of its
arguments. We further assume that f is convex.
The costs associated with increased amounts of BOD and reduced amounts of DO are much
more difﬁcult to quantify since the stream is used by many institutions for a variety of purposes
(e.g., agricultural, industrial, municipal, recreational), and the disutility caused by a decrease in
the water quality varies with the user. Therefore, instead of attempting to quantify these costs let
us suppose that some minimum water quality standards are set. Let q be the minimum acceptable
DO concentration and let ¯ z be the maximum permissible BOD concentration. Then we face the
5.2. DUALITY THEORY 69
following NP:
Maximize −
N
¸
i=1
T
¸
t=1
f
i

i
(t), s
i
(t), m
i
(t))
subject to (5.32), (5.33), and
−q
i
(t) ≤ −q , i = 1, . . . , N; t = 1, . . . , T,
z
i
(t) ≤ ¯ z , i = 1, . . . , N; t = 1, . . . , T,
0 ≤ π
i
(t) ≤ 1 , i = 1, . . . , N; t = 1, . . . , T.
(5.34)
Suppose that all the treatment facilities are in the control of a single public agency. Then assuming
that the agency is required to maintain the standards (q, ¯ z) and it does this at a minimum cost it will
solve the NP (5.34) and arrive at an optimal solution. Let the minimum cost be m(q, ¯ z). But if
there is no such centralized agency, then the individual polluters may not (and usually do not) have
any incentive to cooperate among themselves to achieve these standards. Furthermore, it does not
make sense to enforce legally a minimum standard q
i
(t) ≥ q, z
i
(t) ≤ ¯ z on every polluter since the
pollution levels in the ith area depend upon the pollution levels on all the other areas lying upstream.
On the other hand, it may be economically and politically acceptable to tax individual polluters in
proportion to the amount of pollutants discharged by the individual. The question we now pose
is whether there exist tax rates such that if each individual polluter minimizes its own total cost
(i.e., cost of waste treatment + tax on remaining pollutants), then the resulting water quality will be
acceptable and, furthermore, the resulting amount of waste treatment is carried out at the minimum
expenditure of resources (i.e., will be an optimal solution of (5.34)).
It should be clear from the duality theory that the answer is in the afﬁrmative. To see this let
w
i
(t) = (z
i
(t), −q
i
(t))

, let w(t) = (w
1
(t), . . . , w
N
(t)), and let w = (w(1), . . . , w(t)). Then we
can solve (5.32) and (5.33) for w and obtain
w = b +Ar , (5.35)
where the matrix A and the vector b depend upon the known parameters and initial conditions, and
r is the NT-dimensional vector with components (1 − π
i
(t))s
i
(t)m
i
(t). Note that the coefﬁcients
of the matrix must be non-negative because an increase in any component of r cannot decrease the
BOD levels and cannot increase the DO levels. Using (5.35) we can rewrite (5.34) as follows:
Maximize −
¸
i
¸
t
f
i

i
(t), s
i
(t), m
i
(t))
subject to b +Ar ≤ ¯ w ,
0 ≤ π
i
(t) ≤ 1 , i = 1, . . . , N; t = 1, . . . , T,
(5.36)
where the 2NT-dimensional vector ¯ w has its components equal to −q or ¯ z in the obvious manner.
By the duality theorem there exists a 2NT-dimensional vector λ

≥ 0, and an optimal solution
π

i
(t), i = 1, . . . , N, t = 1, . . . , T, of the problem:
Maximize −
¸
i
¸
t
f
i

i
(t), s
i
(t), m
i
(t)) −λ

(b +Ar −w)
subject to 0 ≤ π
i
(t) ≤ 1, i = 1, . . . , N; t = 1, . . . , T ,
(5.37)
such that ¦π

i
(t)¦ is also an optimal solution of (5.36) and, furthermore, the optimal values of (5.36)
and (5.37) are equal. If we let p

= A

λ

≥ 0, and we write the components of p

as p

i
(t) to match
70 CHAPTER 5. NONLINEAR PROGRAMMING
with the components (1−π
i
(t))s
i
(t)m
i
(t) of r we can see that (5.37) is equivalent to the set of NT
problems:
Maximize −f
i

i
(t), s
i
(t), m
i
(t)) −p

i
(t)(1 −π
i
(t))s
i
(t)m
i
(t)
0 ≤ π
i
(t) ≤ 1 ,
i = 1, . . . , N; t = 1, . . . , T .
(5.38)
Thus, p

i
(t) is optimum tax per mg of BOD in area i during period t.
Before we leave this example let us note that the optimum dual variable or shadow price λ

plays an important role in a larger framework. We noted earlier that the quality standard (q, ¯ z)
was somewhat arbitrary. Now suppose it is proposed to change the standard in the ith area during
period t to q +∆q
i
(t) and ¯ z +∆z
i
(t). If the corresponding components of λ

are λ
q∗
i
(t) and λ
z∗
i
(t),
then the change in the minimum cost necessary to achieve the new standard will be approximately
λ
q∗
i
(t)∆q
i
(t) + λ
z∗
i
(t)∆z
i
(t). This estimate can now serve as a basis in making a beneﬁts/cost
analysis of the proposed new standard.
An important special case of NP is the quadratic programming (QP) problem:
Maximize c

x −
1
2
x

Px
subject to Ax ≤ b, x ≥ 0 ,
(5.39)
where x ∈ R
n
is the decision variable , c ∈ R
n
, b ∈ R
m
are ﬁxed, A is a ﬁxed m n matrix and
P = P

is a ﬁxed positive semi-deﬁnite matrix.
Theorem 1: A vector x

∈ R
n
is optimal for (5.39) iff there exist λ

∈ R
m
, µ

∈ R
n
, such that
Ax

≤ b, x

≥ 0
c −Px

= A

λ

−µ

, λ

≥ 0, µ

≥ 0 ,

)

(Ax

−b) = 0 , (µ

)

x

= 0 .
(5.40)
Proof: By Lemma 3 of 1.3, CQ is satisﬁed, hence the necessity of these conditions follows from
Theorem 2 of 1.2. On the other hand, since P is positive semi-deﬁnite it follows from Exercise 6
of Section 1.2 that f
0
: x → c

x −1/2 x

Px is a concave function, so that the sufﬁciency of these
conditions follows from Theorem 4 of 1.2. ♦
From (5.40) we can see that x

is optimal for (5.39) iff there is a solution (x

, y

, λ

, µ

) to
(5.41), (5.42), and (5.43):
Ax +I
m
Y = b
−Px −A

λ +I
n
µ = −c ,
(5.41)
x ≥ 0 y ≥ 0, λ ≥ 0, µ ≥ 0 , (5.42)
µ

x = 0 , λ

y = 0 . (5.43)
Suppose we try to solve (5.41) and (5.42) by Phase I of the Simplex algorithm (see 4.3.2). Then we
must apply Phase II to the LP:
Maximize −
m
¸
i=1
z
i

n
¸
j=1
ξ
j
5.4. COMPUTATIONAL METHOD 71
subject to
Ax +I
m
y +z = b
−Px −A

λ +I
n
µ +ξ = −c
x ≥ 0, y ≥ 0, λ ≥ 0, µ ≥ 0, z ≥ 0, ξ ≥ 0,
(5.44)
starting with a basic feasible solution z = b, ξ = −c. (We have assumed, without loss of generality,
that b ≥ 0 and −c ≥ 0.) If (5.41) and (5.42) have a solution then the maximum value in (5.44) is 0.
We have the following result.
Lemma 1: If (5.41), (5.42), and (5.43) have a solution, then there is an optimal basic feasible solution
of (5.44) which is also a solution f (5.41), (5.42), and (5.43).
Proof: Let ˆ x, ˆ y,
ˆ
λ, ˆ µ be a solution of (5.41), (5.42), and (5.43). Then ˆ x, ˆ y,
ˆ
λ, ˆ µ, ˆ z = 0,
ˆ
ξ = 0 is
an optimal solution of (5.44). Furthermore, from (5.42) and (5.43) we see that at most (n + m)
components of (ˆ x, ˆ y,
ˆ
λ, ˆ µ) are non-zero. But then a repetition of the proof of Lemma 1 of 4.3.1 will
also prove this lemma. ♦
This lemma suggests that we can apply the Simplex algorithm of 4.3.2 to solve (5.44), starting
with the basic feasible solution z = b, ξ = −c, in order to obtain a solution of (5.41), (5.42), and
(5.43). However, Step 2 of the Simplex algorithm must be modiﬁed as follows to satisfy (5.43):
If a variable x
j
is currently in the basis, do not consider µ
j
as a candidate for entry into the basis;
if a variable y
i
is currently in the basis, do not consider λ
i
as a candidate for entry into the basis. If
it not possible to remove the z
i
and ξ
j
from the basis, stop.
The above algorithm is due to Wolfe [1959]. The behavior of the algorithm is summarized below.
Theorem 2: Suppose P is positive deﬁnite. The algorithm will stop in a ﬁnite number of steps at an
optimal basic feasible solution (ˆ x, ˆ y,
ˆ
λ, ˆ µ, ˆ z,
ˆ
ξ) of (5.44). If ˆ z = 0 and
ˆ
ξ = 0 then (ˆ x, ˆ y,
ˆ
λ, ˆ µ) solve
(5.41), (5.42), and (5.43) and ˆ x is an optimal solution of (5.39). If ˆ z = 0 or
ˆ
ξ = 0, then there is no
solution to (5.41), (5.42), (5.43), and there is no feasible solution of (5.39).
For a proof of this result as well as for a generalization of the algorithm which permits positive
semi-deﬁnite P see (Cannon, Cullum, and Polak [1970], p. 159 ff).
5.4 Computational Method
Maximize f
0
(x)
subject to f
i
(x) ≤ 0, i = 1, . . . , m ,
(5.45)
where x ∈ R
n
, f
i
: R
n
→ R, 0 ≤ i ≤ m, are differentiable. Let Ω ⊂ R
n
denote the set of
feasible solutions. For ˆ x ∈ Ω deﬁne the function ψ(ˆ x) : R
n
→ R by
ψ(ˆ x)(h) = max¦−f
0x
(ˆ x)h, f
1
(ˆ x) +f
1x
(ˆ x)h, . . . , f
m
(ˆ x) +f
mx
(ˆ x)h¦.
Consider the problem:
Minimize ψ(ˆ x)(h)
subject to −ψ(ˆ x)(h) −f
0x
(ˆ x)h ≤ 0 ,
−ψ(ˆ x)(h) +f
i
(ˆ x)f
ix
h ≤ 0 ,
1 ≤ i ≤ m , −1 ≤ h
j
≤ 1 , 1 ≤ j ≤ n .
(5.46)
72 CHAPTER 5. NONLINEAR PROGRAMMING
.
f
0
(x) = F
0
(x

) > f
0
(x
k
)
f
0
(x) = f
0
(x
k
)
f
2
= 0
f
1
= 0

f
3
= 0
f
2
(x
k
)
x
k
f
3
(x
k
)
f
1
(x
k
)
f
0
(x
k
)
h(x
k
)
Figure 5.8: h(x
k
) is a feasible direction.
Call h(ˆ x) an optimum solution of (5.46) and let h
0
(ˆ x) = ψ(ˆ x)(h(ˆ x)) be the minimum value at-
tained. (Note that by Exercise 1 of 4.5.1 (5.46) can be solved as an LP.)
The following algorithm is due to Topkis and Veinott [1967].
Step 1. Find x
0
∈ Ω, set k = 0, and go to Step 2.
Step 2. Solve (5.46) for ˆ x = x
k
and obtain h
0
(x
k
), h(x
k
). If h
0
(x
k
) = 0, stop, otherwise go to Step
3.
Step 3. Compute an optimum solution µ(x
k
) to the one-dimensional problem,
Maximize f
0
(x
k
+µh(x
k
)) ,
subject to (x
k
+µh(x
k
)) ∈ Ω, µ ≥ 0 ,
and go to Step 4.
Step 4. Set x
k+1
= x
k
+µ(x
k
)h(x
k
), set k = k + 1 and return to Step 2.
The performance of the algorithm is summarized below.
Theorem 1: Suppose that the set
Ω(x
0
) = ¦x[x ∈ Ω, f
0
(x) ≥ f
0
(x
0

is compact, and has a non-empty interior, which is dense in Ω(x
0
). Let x

be any limit point of
the sequence x
0
, x
1
, . . . , x
k
, . . . , generated by the algorithm. Then the Kuhn-Tucker conditions are
satisﬁed at x

.
For a proof of this result and for more efﬁcient algorithms the reader is referred to (Polak [1971]).
Remark: If h
0
(x
k
) < 0 in Step 2, then the direction h(x
k
) satisﬁes f
0x
(x
k
)h(x
k
) > 0, and f
i
(x
k
)+
f
ix
(x
K
)h(x
k
) < 0, 1 ≤ i ≤ m. For this reason h(x
k
) is called a (desirable) feasible direction.
(See Figure 5.8.)
5.5. APPENDIX 73
5.5 Appendix
The proofs of Lemmas 4,7 of Section 2 are based on the following extremely important theorem
(see Rockafeller [1970]).
Separation theorem for convex sets. Let F, Gbe convex subsets of R
n
such that the relative interiors
of F, G are disjoint. Then there exists λ ∈ R
n
, λ = 0, and θ ∈ R such that
λ

g ≤ θ for all g ∈ G
λ

f ≥ θ for all f ∈ F .
Proof of Lemma 4: Since M is stable at
ˆ
b there exists K such that
M(b) −M(
ˆ
b) ≤ K[b −
ˆ
b[ for all b ∈ B . (5.47)
In R
1+m
consider the sets
F = ¦(r, b)[b ∈ R
m
, r > K[b −
ˆ
b[¦ ,
G = ¦(r, b)[b ∈ B, r ≤ M(b) −M(
ˆ
b)¦ .
It is easy to check that F, G are convex, and (5.47) implies that F ∩ G = φ. Hence, there exist

0
, . . . , λ
m
) = 0, and θ such that
λ
0
r +
m
¸
i=1
λ
i
b
i
≤ θ for (r, b) ∈ G ,
λ
0
r +
m
¸
i=1
λ
i
b
i
≥ θ for (r, b) ∈ F .
(5.48)
From the deﬁnition of F, and the fact that (λ
0
, . . . , λ
m
) = 0, it can be veriﬁed that (5.49) can hold
only if λ
0
> 0. Also from (5.49) we can see that
m
¸
i=1
λ
i
ˆ
b
i
≥ θ, whereas from (5.48)
m
¸
i=1
λ
i
ˆ
b
i
≤ θ,
so that
m
¸
i=1
λ
i
ˆ
b
i
= θ. But then from (5.48) we get
M(b) −M(
ˆ
b) ≤
1
λ
0
[θ −
m
¸
i=1
λ
i
b
i
] =
m
¸
i=1
(−
λ
i
λ
0
)(b
i

ˆ
b). ♦
Proof of Lemma 7: Since
ˆ
b is in the interior of B, there exists ε > 0 such that
b ∈ B whenever [b −
ˆ
b[ < ε . (5.49)
In R
1+m
consider the sets
F = ¦(r,
ˆ
b)[r > M(
ˆ

G = ¦(r, b)[b ∈ B, r ≤ M(b)¦ .
Evidently, F, G are convex and F ∩ G = φ, so that there exist (λ
0
, . . . , λ
m
) = 0, and θ such that
λ
0
r +
m
¸
i=1
λ
i
ˆ
b
i
≥ θ , for r > M(
ˆ
b) , (5.50)
74 CHAPTER 5. NONLINEAR PROGRAMMING
λ
0
r +
m
¸
i=1
λ
i
ˆ
b
i
≤ θ , for (r, b) ∈ G . (5.51)
From (5.49), and the fact that (λ
0
, . . . , λ
m
) = 0 we can see that (5.50) and (5.51) imply λ
0
> 0.
From (5.50),(5.51) we get
λ
0
M(
ˆ
b) +
m
¸
i=1
λ
i
ˆ
b
i
= θ ,
so that (5.52) implies
M(b) ≤ (
ˆ
b) +
m
¸
i=1
(−
λ
i
λ
0
)(b
i

ˆ
b
i
) . ♦
Chapter 6
SEQUENTIAL DECISION PROBLEMS:
DISCRETE-TIME OPTIMAL
CONTROL
In this chapter we apply the results of the last two chapters to situations where decisions have to be
made sequentially over time. A very important class of problems where such situations arise is in
the control of dynamical systems. In the ﬁrst section we give two examples, and in Section 2 we
derive the main result.
6.1 Examples
The trajectory of a vertical sounding rocket is controlled by adjusting the rate of fuel ejection which
generates the thrust force. Speciﬁcally suppose that the equations of motion are given by (6.1).
˙ x
1
(t) = x
2
(t)
˙ x
2
(t) = −
C
D
x
3
(t)
ρ(x
1
(t))x
2
2
(t) −g +
C
T
x
3
(t)
u(t)
˙ x
3
(t) = −u(t) ,
(6.1)
where x
1
(t) is the height of the rocket from the ground at time t, x
2
(t) is the (vertical) speed at
time t, x
3
(t) is the weight of the rocket (= weight of remaining fuel) at time t. The “dot” denotes
differentiation with respect to t. These equations can be derived from the force equations under the
assumption that there are four forces acting on the rocket, namely: inertia = x
3
¨ x
1
= x
3
˙ x
2
; drag
force = C
D
ρ(x
1
)x
2
2
where C
D
is constant, ρ(x
1
) is a friction coefﬁcient depending on atmospheric
density which is a function of x
1
; gravitational force = gx
3
with g assumed constant; and thrust
force C
T
˙ x
3
, assumed proportional to rate of fuel ejection. See Figure 6.1. The decision variable at
time t is u(t), the rate of fuel ejection. At time 0 we assume that (x
1
(0), x
2
(0), x
3
(0)) = (0, 0, M);
that is, the rocket is on the ground, at rest, with initial fuel of weight M. At a prescribed ﬁnal time
t
f
, it is desired that the rocket be at a position as high above the ground as possible. Thus, the
75
76 CHAPTER 6. DISCRETE-TIME OPTIMAL CONTROL
decision problem can be formalized as (6.2).
Maximize x
1
(t
f
)
subject to ˙ x(t) = f(x(t), u(t)), 0 ≤ t ≤ t
f
x(0) = (0, 0, M)
u(t) ≥ 0, x
3
(t) ≥ 0, 0 ≤ t ≤ t
f
,
(6.2)
where x = (x
1
, x
2
, x
3
)

, f : R
3+1
→ R
3
is the right-hand side of (6.1). The constraint inequalities
u(t) ≥ 0 and x
3
(t) ≥ 0 are obvious physical constraints.
x
3
¨ x
1
= inertia
C
D
ϕ(x
1
)x
2
2
= drag
gx
3
= gravitational force
C
R
˙ x
3
= thrust
Figure 6.1: Forces acting on the rocket.
The decision problem (6.2) differs from those considered so far in that the decision variables,
which are functions u : [0, t
f
] → R, cannot be represented as vectors in a ﬁnite-dimensional
space. We shall treat such problems in great generality in the succeeding chapters. For the moment
we assume that for computational or practical reasons it is necessary to approximate or restrict
the permissible function u() to be constant over the intervals [0, t
1
), [t
1
, t
2
), . . . , [t
N−1
, t
f
), where
t
1
, t
2
, . . . , t
N−1
are ﬁxed a priori. But then if we let u(i) be the constant value of u() over [t
i
, t
i+1
),
we can reformulate (6.2) as (6.3):
Maximize x
1
(t
N
)(t
N
= t
f
)
subject to x(t
i+1
) = g(i, x(t
i
), u(i)), i = 0, 1, . . . , N −1
x(t
0
) = x(0) = (0, 0, M)
u(i) ≥ 0, x
3
(t
i
) ≥ 0, i = 0, 1, . . . , N .
(6.3)
In (6.3) g(i, x(t
1
), u(i)) is the state of the rocket at time t
i+1
when it is in state x(t
i
) at time t
i
and
u(t) ≡ u(i) for t
i
≤ t < t
i+1
.
As another example consider a simple inventory problem where time enters discretely in a natural
fashion. The Squeezme Toothpaste Company wants to plan its production and inventory schedule
for the coming month. It is assumed that the demand on the ith day, 0 ≤ i ≤ 30, is d
1
(i) for
6.2. MAIN RESULT 77
their orange brand and d
2
(i) for their green brand. To meet unexpected demand it is necessary that
the inventory stock of either brand should not fall below s > 0. If we let s(i) = (s
1
(i), s
2
(i))

denote the stock at the beginning of the ith day, and m(i) = (m
1
(i), m
2
(i))

denote the amounts
manufactured on the ith day, then clearly
s(i + 1) +s(i) +m(i) −d(i) ,
where d(i) = (d
1
(i), d
2
(i))

. Suppose that the initial stock is ˆ s, and the cost of storing inventory s
for one day is c(s) whereas the cost of manufacturing amount mis b(m). The the cost-minimization
decision problem can be formalized as (6.4):
Maximize
30
¸
i=0
(c(s(i)) +b(m(i)))
subject to s(i + 1) = s(i) +m(i) −d(i), 0 ≤ i ≤ 29
s(0) = ˆ s
s(i) ≥ (s, s)

, m(i) ≥ 0, 0 ≤ i ≤ 30 .
(6.4)
Before we formulate the general problem let us note that (6.3) and (6.4) are in the form of non-
linear programming problems. The reason for treating these problems separately is because of their
practical importance, and because the conditions of optimality take on a special form.
6.2 Main Result
The general problem we consider is of the form (6.5).
Maximize
N−1
¸
i=0
f
0
(i, x(i), u(i))
subject to
dynamics : x(i + 1) −x(i) = f(i, x(i), u(i)), i = 0, . . . , N −1 ,
initial condition: q
0
(x(0) ≤ 0, g
0
(x(0)) = 0 ,
ﬁnal condition: q
N
(x(N)) ≤ 0, g
N
(x(N)) = 0 ,
state-space constraint: q
i
(x(i)) ≤ 0, i = 1, . . . , N −1 ,
control constraint: h
i
(u(i)) ≤ 0, i = 0, . . . , N −1 .
(6.5)
Here x(i) ∈ R
n
, u(i) ∈ R
p
, f
0
(i, , ) : R
n+p
→ R, f(i, , ) : R
n+p
→ R
n
, q
i
: R
n

R
m
i
, g
i
: R
n
→ R

i
, h
i
: R
p
→ R
s
i
are given differentiable functions. We follow the control
theory terminology, and refer to x(i) as the state of the system at time i, and u(i) as the control or
input at time i.
We use the formulation mentioned in the Remark following Theorem 3 of V.1.2, and construct
the Lagrangian function L by
L(x(0), . . . , x(N); u(0), . . . , u(N −1); p(1), . . . , p(N);
λ
0
, . . . , λ
N
; α
0
, α
N
; γ
0
, . . . , γ
N−1
)
78 CHAPTER 6. DISCRETE-TIME OPTIMAL CONTROL
=
N−1
¸
i=0
f
0
(i, x(i), u(i)) −

N−1
¸
i=0
(p(i + 1))

(x(i + 1) −x(i) −f(i, x(i), u(i)))+
N
¸
i=0

i
)

q
i
(x(i)) + (α
0
)

g
0
(x(0)) + (α
N
)

g
N
(x(N)) +
N−1
¸
i=0

i
)

h
i
(u(i))
¸
.
Suppose that CQ is satisﬁed for (6.5), and x

(0), . . . , x

(N); u

(0), . . . , u

(N −1), is an optimal
solution. Then by Theorem 2 of 5.1.2, there exist p

(i) in R
n
for 1 ≤ i ≤ N, λ
i∗
≥ 0 in R
m
i
for
0 ≤ i ≤ N, α
i∗
in R

i
for i = 0, N, and γ
i∗
≥ 0 in R
s
i
for 0 ≤ i ≤ N −1, such that
(A) the derivative of L evaluated at these points vanishes,
and
(B) λ
i∗
q
i
(x

(i)) = 0 for 0 ≤ i ≤ N, γ
i∗
h
i
(u

(i)) = 0 for 0 ≤ i ≤ N −1 .
We explore condition (A) by taking various partial derivatives.
Differentiating L with respect to x(0) gives
f
0x
(0, x

(0), u

(0)) −¦−(p

(1))

−(p

(1))

[f
x
(0, x

(0), u

(0))]
+(λ
0∗
)

[q
0x
(x

(0))] + (α
0∗
)

[g
0x
(x

(0))]¦ = 0 ,
or
p

(0) −p

(1) = [f
x
(0, x

(0), u

(x))]

p

(1)
+[f
0x
(0, x

(0), u

(0))]

−[q
0x
(x

(0))]

λ
0∗
,
(6.6)
where we have deﬁned
p

(0) = [g
0x
(x

(x))]

α
0∗
. (6.7)
Differentiating L with respect to x(i), 1 ≤ i ≤ N −1, and re-arranging terms gives
p

(i) −p

(i + 1) = [f
x
(i, x

(i), u

(i))]

p

(i + 1)
+[f
0x
(i, x

(i), u

(i))]

−[q
ix
(x

(i))]

λ
i∗
.
(6.8)
Differentiating L with respect to x(N) gives,
p

(N) = −[g
Nx
(x

(N))]

α
N∗
−[q
Nx
(x

(N))]

λ
N∗
.
It is convenient to replace α
N∗
by −α
N∗
so that the equation above becomes (6.9)
p

(N) = [g
Nx
(x

(N))]

α
N∗
−[q
Nx
(x

(N))]

λ
N∗
. (6.9)
Differentiating L with respect to u(i), 0 ≤ i ≤ N −1 gives
[f
0u
(i, x

(i), u

(i))]

+ [f
u
(i, x

(i), u

(i))]

p

(i +l) −[h
iu
(u

(i))]

γ
i∗
= 0 . (6.10)
We summarize our results in a convenient form in
Table 6.1
Remark 1: Considerable elegance and mnemonic simpliﬁcation is achieved if we deﬁne the
Hamiltonian function H by
6
.
2
.
M
A
I
N
R
E
S
U
L
T
7
9
Suppose x

(0), . . . , x

(N);
u

(0), . . . , u

(N −1) maximizes
N−1
¸
i=0
f
0
(i, x(i), u(i)) subject
to the constraints below
then there exist p

(N); λ
0∗
, . . . , λN

; α
0∗
, α
N∗;
γ
0∗
, . . . , γ
N−1

, such that
dynamics: i = 0, . . . , N −1
x(i + 1) −x(i) = f(i, x(i), u(i))
initial condition:
q
0
(x

(0)) ≤ 0, g
0
(x

(0)) = 0
ﬁnal conditions:
q
N
(x

(N)) ≤ 0, g
N
(x

(N)) = 0
state space constraint:
i = 1, . . . , N −1
q
i
(x

(i)) ≤ 0
control constraint:
i = 0, . . . , N −1
h
i
(u

(i)) ≤ 0
adjoint equations: i = 0, . . . , N −1
p

(i) −p

(i + 1) = [f
x
(i, x

(i), u

(i)]

p

(i + 1)
+[f
0x
(i, x

(i), u

(i)]

−[q
ix
(x

(i)]

γ
i∗
transversality conditions:
p

(0) = [g
0x
(x

(0))]

α
0∗
p

(N) = [g
Nx
(x

(N))]

α
N∗
−[q
Nx
(x

(N))]

λ
N∗
[f
0u
(i, x

(i), u

(i))]

+ [f
u
(i, x

(i)u

(i))]

.
p

(i
1
) = [h
iu
(u

(i))]

γ
i∗
λ
0∗
≥ 0,

0∗
)

q
0
(x

(0)) = 0
λ
N∗
≥ 0,

N∗)

q
N
(x

(N)) = 0
λ
i∗
≥ 0,

i∗
)

q
i
(x

(i)) = 0
γ
i∗
≥ 0

i∗)

h
i
(u

(i) = 0
T
a
b
l
e
6
.
1
:
80 CHAPTER 6. DISCRETE-TIME OPTIMAL CONTROL
H(i, x, u, p) = f
0
(i, x, u) +p

f(i, x, u) .
The dynamic equations then become
x

(i + 1) −x

(i) = [H
p
(i, x

(i), u

(i), p

(i + 1))]

,
0 ≤ i ≤ N −1 .
(6.11)
and the adjoint equations (6.6) and (6.8) become
p

(i) −p

(i + 1) = [H
x
(i, x

(i), u

(i), u

(i), p

(i + 1))]

−[q
ix
(x

(i))]

λ
i∗
,
0 ≤ i ≤ N −1 ,
whereas (6.10) becomes
[h
iu
(u

(i))]

γ
i∗
= [H
u
(i, x

(i), u

(i), p

(i + 1))]

, 0 ≤ i ≤ N −1 . (6.12)
Remark 2: If we linearize the dynamic equations about the optimal solution we obtain
δx(i + 1) −δx(i) = [f
x
(i, x

(i), u

(i))]δx(i) + [f
u
(i, x

(i, x

, (i), u

(i))]δu(i) ,
whose homogeneous part is
z(i + 1) −z(i) = [f
x
(i, x

(i), u

(i))]z(i) ,
which has for it adjoint the system
r(i) −r(i + 1) = [f
x
(i, x

(i), u

(i))]

r(i + 1) . (6.13)
Since the homogeneous part of the linear difference equations (6.6), (6.8) is (6.13), we call (6.6),
(6.8) the adjoint equations, and the p

Remark 3: If the f
0
(i, , ) are concave and the remaining function in (6.5) are linear, then CQ is
satisﬁed, and the necessary conditions of Table 6.1 are also sufﬁcient. Furthermore, in this case we
see from (6.13) that u

(i) is an optimal solution of
Maximize H(i, x

(i), u, p

(i + 1)),
subject to h
i
(u) ≤ 0 .
For this reason the result is sometimes called the maximum principle.
Remark 4: The conditions (6.7), (6.9) are called transversality conditions for the following reason.
Suppose q
0
≡ 0, q
N
≡ 0, so that the initial and ﬁnal conditions read g
0
(x(0)) = 0, g
N
(x(N)) = 0,
which describe surfaces in R
n
. Conditions (6.7), (6.9) become respectively p

(0) = [g
0x
(x

(0))]

α
0∗
, p

(N) =
[g
Nx
(x(N))]

α
N∗
which means that p

(0) and p

(N) are respectively orthogonal or transversal to
the initial and ﬁnal surfaces. Furthermore, we note that in this case the initial and ﬁnal conditions
specify (
0
+
n
) conditions whereas the transversality conditions specify (n−
0
)+(n−
n
) condi-
tions. Thus, we have a total of 2n boundary conditions for the 2n-dimensional system of difference
equations (6.5), (6.12); but note that these 2n boundary conditions are mixed, i.e., some of them
refer to the initial time 0 and the rest refer to the ﬁnal time.
6.2. MAIN RESULT 81
Exercise 1: For the regulator problem,
Maximize
1
2
N−1
¸
i=0
x(i)

Qx(i) +
1
2
N−1
¸
i=0
u(i)

Pu(i)
subject to x(i + 1) −x(i) = Ax(i) +Bu(i), 0 ≤ i ≤ N −1
x(0) = ˆ x(0),
u(i) ∈ R
p
, 0 ≤ i ≤ N −1 ,
where x(i) ∈ R
n
, A and B are constant matrices, ˆ x(0) is ﬁxed, Q = Q

is positive semi-deﬁnite,
and P = P

is positive deﬁnite, show that the optimal solution is unique and can be obtained by
solving a 2n-dimensional linear difference equation with mixed boundary conditions.
Exercise 2: Show that the minimal fuel problem,
Minimize
N−1
¸
i=0

¸
P
¸
j=1
[(u(i))
j
[
¸

,
subject to x(i + 1) −x(i) = Ax(i) +Bu(i), 0 ≤ i ≤ N −1
x(0) = ˆ x(0), x(N) = ˆ x(N) ,
u(i) ∈ R
p
, [u(i))
j
[ ≤ 1, 1 ≤ j ≤ p, 0 ≤ i ≤ N −1
can be transformed into a linear programming problem. Here ˆ x(0), ˆ x(N) are ﬁxed, A and B are as
in Exercise 1.
82 CHAPTER 6. DISCRETE-TIME OPTIMAL CONTROL
Chapter 7
SEQUENTIAL DECISION PROBLEMS:
CONTINUOUS-TIME OPTIMAL
CONTROL OF LINEAR SYSTEMS
We will investigate decision problems similar to those studied in the last chapter with one (math-
ematically) crucial difference. A choice of control has to be made at each instant of time t where
t varies continuously over a ﬁnite interval. The evolution in time of the state of the systems to be
controlled is governed by a differential equation of the form:
˙ x(t) = f(t, x(t), u(t)) ,
where x(t) ∈ R
n
and u(t) ∈ R
p
are respectively the state and control of the system at time t.
To understand the main ideas and techniques of analysis it will prove proﬁtable to study the linear
case ﬁrst. The general nonlinear case is deferred to the next chapter. In Section 1 we present the
general linear problem and study the case where the initial and ﬁnal conditions are particularly
simple. In Section 2 we study more general boundary conditions.
7.1 The Linear Optimal Control Problem
We consider a dynamical system governed by the linear differential equation (7.1):
˙ x(t) = A(t)x(t) +B(t)u(t), t ≥ t
0
. (7.1)
Here A() and B() are n n- and n p-matrix valued functions of time; we assume that they are
piecewise continuous functions. The control u() is constrained to take values in a ﬁxed set Ω ⊂ R
p
,
and to be piecewise continuous.
Deﬁnition: A piecewise continuous function u : [t
0
, ∞) → Ω will be called an admissible control.
| denotes the set of all admissible controls.
Let c ∈ R
n
, x
0
∈ R
n
be ﬁxed and let t
f
≥ t
0
be a ﬁxed time. We are concerned with the
83
84 CHAPTER 7. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL
decision problem (7.2).
Maximize c

x(t
f
),
subject to
dynamics: ˙ x(t) = A(t)x(t) +B(t)u(t) , t
0
≤ t ≤ t
f
,
initial condition: x(t
0
) = x
0
,
ﬁnal condition: x(t
f
) ∈ R
n
,
control constraint: u() ∈ | .
(7.2)
Deﬁnition: (i) For any piecewise continuous function u() : [t
0
, t
f
] → R
p
, for any z ∈ R
n
, and
any t
0
≤ t
1
≤ t
2
≤ t
f
let
φ(t
2
, t
1
, z, u)
denote the state of (7.1) at time t
2
, if a time t
1
it is in state z, and the control u() is applied.
(ii) Let
K(t
2
, t
1
, z) = ¦φ(t
2
, t
1
, z, u)[u ∈ |¦ .
Thus, K(t
2
, t
1
, z) is the set of states reachable at time t
2
starting at time t
1
in state z and using
admissible controls. We call K the reachable set.
Deﬁnition: Let Φ(t, τ), t
0
≤ τ ≤ t ≤ t
f
, be the transition-matrix function of the homogeneous
part of (7.1), i.e., Φ satisﬁes the differential equation
∂Φ
∂t
(t, τ) = A(t)Φ(t, τ) ,
and the boundary condition
Φ(t, t) ≡ I
n
.
The next result is well-known. (See Desoer [1970].)
Lemma 1: φ(t
2
, t
1
, z, u) = Φ(t
2
, t
1
)z +

t
2
t
1
Φ(t
2
, τ)B(τ)u(τ)dτ.
Exercise 1: (i) Assuming that Ω is convex, show that | is a convex set. (ii) Assuming that | is
convex show that K(t
2
, t
1
, z) is a convex set. (It is a deep result that K(t
2
, t
1
, z) is convex even if
Ω is not convex (see Neustadt [1963]), provided we include in | any measurable function
u : [t
0
, ∞) → Ω.)
Deﬁnition: Let K ⊂ R
n
, and let x

∈ K. We say that c is the outward normal to a hyperplane
supporting K at x

if c = 0, and
c

x

≥ c

x for all x ∈ K .
The next result gives a geometric characterization of the optimal solutions of (2).
Lemma 2: Suppose c = 0. Let u

() ∈ | and let x

(t) = φ(t, t
0
, x
0
, u

). Then u

is an optimal
solution of (2) iff
(i) x

(t
f
) is on the boundary of K = K(t
f
, t
0
, x
0
), and
(ii) c is the outward normal to a hyperplane supporting K at x

. (See Figure 7.1.)
Proof: Clearly (i) is implied by (ii) because if x

(t
f
) is in the interior of K there is δ > 0 such
that (x

(t
f
) +δc) ∈ K; but then
7.1. THE LINEAR OPTIMAL CONTROL PROBLEM 85
x
3
c
x
2
x
1
c
x

(t
f
)
π

= ¦x[c

x = c

x

(t
f

K
Figure 7.1: c is the outward normal to π

supporting K at x

(t
f
)
.
c

(x

(t
f
) +δc) = c

x

(t
f
) +δ[c[
2
> c

x

(t
f
) .
Finally, from the deﬁnition of K it follows immediately that u

is optimal iff c

x

(t
f
) ≥ c

x for all
x ∈ K . ♦
The result above characterizes the optimal control u

in terms of the ﬁnal state x

(t
f
). The beauty
and utility of the theory lies in the following result which translates this characterization directly in
terms of u

.
Theorem 1: Let u

() ∈ | and let x

(t) = φ(t, t
0
, x
0
, u

), t
0
≤ t ≤ t
f
. Let p

(t) be the solution
of (7.3) and (7.4):

(t) = −A

(t)p

(t) , t
0
≤ t ≤ t
f
. (7.3)
ﬁnal condition: p

(t
f
) = c . (7.4)
Then u

() is optimal iff
(p

(t))

B(t)u

(t) = sup¦(p

(t))

B(t)v[v ∈ Ω¦ , (7.5)
for all t ∈ [t
0
, t
f
], except possibly for a ﬁnite set.
Proof: u

() is optimal iff for every u() ∈ |
(p

(t
f
))

[Φ(t
f
, t
0
)x
0
+

t
f
t
0
Φ(t
f
, τ)B(τ)u

(τ)dτ]
≥ (p

(t
f
))

[Φ(t
f
, t
0
)x
0
+

t
f
t
0
Φ(t
f
, τ)B(τ)u(τ)dτ] ,
which is equivalent to (7.6).

t
f
t
0
(p

(t
f
))

Φ(t
f
, τ)B(τ)u

(τ)dτ

t
f
t
0
(p

(t
f
))

Φ(t
f
, τ)B(τ)u(τ)dτ
(7.6)
86 CHAPTER 7. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL
Now by properties of the adjoint equation we know that p

(t))

= (p

(t
f
))

Φ(t
f
, t) so that (7.6) is
equivalent to (7.7),

t
f
t
0
(p

(τ))

B(τ)u

(τ)dτ ≥

t
f
t
0
(p

(τ))

B(τ)u(τ)dτ, (7.7)
and the sufﬁciency of (7.5) is immediate.
To prove the necessity let D be the ﬁnite set of points where the function B() or u

() is discon-
tinuous. We shall show that if u

() is optimal then (7.5) is satisﬁed for t ∈ D. Indeed if this is not
the case, then there exists t

∈ [t
0
, t
f
], t

∈ D, and v ∈ Ω such that
(p

(t

))

B(t

)u

(t

) < (p

(t

))

B(t

)v ,
and since t

is a point of continuity of B() and u

(), it follows that there exists δ > 0 such that
(p

(t))

B(t)u

(t) < (p

(t))

B(t)v, for [t −t

[ < δ . (7.8)
Deﬁne ˜ u() ∈ | by
˜ u(t) =

v [t −t

[ < δ, t ∈ [t
0
, t
f
]
u

(t) otherwise .
Then (7.8) implies that

t
f
t
0
(p

(t))

B(t)˜ u(t)dt >

t
f
t
0
(p

(t))

B(t)u

(t)dt .
But then from (7.7) we see that u

() cannot be optimal, giving a contradiction. ♦
Corollary 1: For t
0
≤ t
1
≤ t
2
≤ t
f
,
(p

(t
2
))x

(t
2
) ≥ (p

(t
2
))

x for all x ∈ K(t
2
, t
1
, x

(t
1
)). (7.9)
Exercise 2: Prove Corollary 1.
Remark 1: The geometric meaning of (7.9) is the following. Taking t
1
= t
0
in (7.9), we see that if
u

() is optimal, i.e., if c = p

(t
f
) is the outward normal to a hyperplane supporting K(t
f
, t
0
, x
0
)
at x

(t
f
), then x

(t) is on the boundary of K(t, t
0
, x
0
) and p

(t) is the normal to a hyperplane
supporting K(t, t
0
, x
0
) at x

(t). This normal is obtained by transporting backwards in time, via
the adjoint differential equation, the outward normal p

(t
f
) at time t
f
. The situation is illustrated
in Figure 7.2.
Remark 2: If we deﬁne the Hamiltonian function H by
H(t, x, u, p) = p

(A(t)x +B(t)u) ,
and we deﬁne M by
M(t, x, p) = sup¦H(t, x, u, p)[u ∈ Ω¦,
then (7.5) can be rewritten as
H(t, x

(t), u

(t), p

(t)) = M(t, x

(t), p

(t)) . (7.10)
This condition is known as the maximum principle.
7.2. MORE GENERAL BOUNDARY CONDITIONS 87
Exercise 3: (i) Show that m(t) = M(t, x

(t), p

(t)) is a Lipschitz function of t. (ii) If A(t), B(t)
are constant, show that m(t) is constant. (Hint: Show that (dm/dt) ≡ 0.)
The next two exercises show how we can obtain important qualitative properties of an optimal
control.
Exercise 4: Suppose that Ω is bounded and closed. Show that there exists an optimal control u

()
such that u

(t) belongs to the boundary of Ω for all t.
Exercise 5: Suppose Ω = [α, β], so that B(t) is an n 1 matrix. Suppose that A(t) ≡ A and
B(t) ≡ B are constant matrices and A has n real eigenvalues. Show that there is an optimal
control u

() and t
0
≤ t
1
≤ t
2
≤ . . . ≤ t
n
≤ t
f
such that u

(t) ≡ α or β on [t
i
, t
i+1
), 0 ≤ i ≤ n.
(Hint: ﬁrst show that (p

(t))

B = γ
1
exp(δ
1
t) +. . . +γ
n
exp(δ
n
(t)) for some γ
i
, δ
i
in R.)
Exercise 6: Assume that K(t
f
, t
0
, x
0
) is convex (see remark in Exercise 1 above). Let
f
0
: R
n
→ R be a differentiable function and suppose that the objective function in (7.2) is
f
0
(x(t
f

x(t
f
). Suppose u

() is an optimal control. Show that u

() satisﬁes the
maximum principle (7.10) where p

() is the solution of the adjoint equation (7.3) with the ﬁnal
condition
p

(t
f
) = f
0
(x

(t
f
)) .
Also show that this condition is sufﬁcient for optimality if f
0
is concave. (Hint: Use Lemma 1 of
5.1.1 to show that if u

() is optimal, then f
0x
(x

(t
f
)(x

(t
f
) −x) ≤ for all x ∈ K(t
f
, t
0
, x
0
).)
7.2 More General Boundary Conditions
We consider the following generalization of (7.2). The notion of the previous section is retained.
Maximize c

x(t
f
)
subject to
dynamics: ˙ x(t) = A(t)x(t) +B(t)u(t), t
0
≤ t ≤ t
f
,
initial condition: G
0
x(t
0
) = b
0
,
ﬁnal condition: G
f
x(t
f
) = b
f
,
control constraint: u() ∈ |, i.e., ¯ : [.

, .
{
] → ⊗ and
u()piecewise continuous.
(7.11)
In (7.11) G
0
and G
f
are ﬁxed matrices of dimensions
0
xn and
f
n respectively, while b
0

R

0
, b
f
∈ R

f
are ﬁxed vectors.
We will analyze the problem in the same way as before. That is, we ﬁrst characterize optimality
in terms of the state at the ﬁnal time, and then translate these conditions in terms of the control. For
convenience let
T
0
= ¦z ∈ R
n
[G
0
z = b
0
¦ ,
T
f
= ¦z ∈ R
n
[G
f
z = b
f
¦ .
Deﬁnition: Let p ∈ R
n
. Let z

∈ T
0
. We say that p is orthogonal to T
0
at z

and we write
p ⊥ T
0
(z

) if
88 CHAPTER 7. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL
R
n
x
0
=
x

(
t
0
)
=
K
(
t
0
,
t
0
,
x
0
) R
n
p

(
t
1
)
x

(
t
1
)
K
(
t
1
,
t
0
,
x
0
)
R
n
p

(
t
2
)
x

(
t
2
)
K
(
t
2
,
t
0
,
x
0
)
R
n
p

(
t
f
)
=
c
x

(
t
f
)
K
(
t
f
,
t
0
,
x
0
)
t
t
f
t
2
t
1
t
0
Figure 7.2: Illustration of (7.9) for t
1
= t
0
.
7.2. MORE GENERAL BOUNDARY CONDITIONS 89
p

(z −z

) = 0 for all z ∈ T
0
.
Similarly if z

∈ T
f
, p ⊥ T
f
(z

) if
p

(z −z

) = 0 for all z ∈ T
f
.
Deﬁnition: Let X(t
f
) = ¦Φ(t
f
, t
0
)z +w[z ∈ T
0
, w ∈ K(t
f
, t
0
, 0)¦.
Exercise 1: X(t
f
) = ¦Φ(t
f
, t
0
, z, u)[z ∈ T
0
, u() ∈ |¦.
Lemma 1: Let x

(t
0
) ∈ T
0
and u

() ∈ |. Let x

(t) = φ(t, t
0
, x

(t
0
), u

), and suppose that
x

(t
f
) ∈ T
f
.
(i) Suppose the Ω is convex. If u

() is optimal, there exist ˆ p
0
∈ R, ˆ p
0
≥ 0 and ˆ p ∈ R
n
, not both
zero, such that
(ˆ p
0
c + ˆ p)

x

(t
f
) ≥ (ˆ p
0
c + ˆ p)

x for all x ∈ X(t
f
) , (7.12)
ˆ p ⊥ T
f
(x

(t
f
)) , (7.13)
[Φ(t
f
, t
0
)]

(ˆ p
0
c + ˆ p) ⊥ T
0
(x

(t
0
)) . (7.14)
(ii) Conversely if there exist ˆ p
0
> 0, and ˆ p such that (7.12) and (7.13) are satisﬁed, then u

() is
optimal and (7.14) is also satisﬁed.
Proof: Clearly u

() is optimal iff
c

x

(t
f
) ≥ c

x for all x ∈ X(t
f
) ∩ T
f
. (7.15)
(i) Suppose that u

() is optimal. In R
1+m
deﬁne sets S
1
, S
2
by
S
1
= ¦(r, x)[r > c

x

(t
f
), x ∈ T
f
¦ , (7.16)
S
2
= ¦(r, x)[r = c

x , x ∈ X(t
f
)¦ . (7.17)
First of all S
1
∩S
2
= φ because otherwise there exists x ∈ X(t
f
) ∩T
f
such that c

x > c

x

(t
f
)

() by (7.15).
Secondly, S
1
is convex since T
f
is convex. Since Ω is convex by hypothesis it follows by Exercise
1 of Section 1 that S
2
is convex.
But then by the separation theorem for convex sets (see 5.5) there exists ˆ p
0
∈ R, ˆ p ∈ R
n
, not
both zero, such that
ˆ p
0
r
1
+ ˆ p

x
1
≥ ˆ p
0
r
2
+ ˆ p

x
2
for all (r
i
, x
i
) ∈ S
i
, i = 1, 2. (7.18)
In particular (7.18) implies that
ˆ p
0
r + ˆ p

x

(t
f
) ≥ ˆ p
0
c

x + ˆ p

x for all x ∈ X(t
f
), r > c

x

(t
f
). (7.19)
Letting r → ∞ we conclude that (7.19) can hold only if ˆ p
0
≥ 0. On the other hand letting r →
c

x

(t
f
) we see that (7.19) can hold only if
ˆ p
0
c

x

(t
f
) + ˆ p

x

(t
f
) ≥ ˆ p
0
c

x + ˆ p

x for all x ∈ X(t
f
) , (7.20)
which is the same as (7.12). Also from (7.18) we get
90 CHAPTER 7. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL
ˆ p
0
r + ˆ p/x ≥ ˆ p
0
c

x

(t
f
) + ˆ p

x

(t
f
) for all r > c

x

(t
f
), x ∈ T
f
,
which can hold only if
ˆ p
1
c

x

(t
f
) + ˆ p

x ≥ ˆ p
0
c

x

(t
f
) + ˆ p

x

(t
f
) for all x ∈ T
f
,
or
ˆ p

(x −x

(t
f
)) ≥ 0 for all x ∈ T
f
(7.21)
But ¦x −x

(t
f
)[x ∈ T
f
¦ = ¦z[G
f
z = 0¦ is a subspace of R
n
, so that (7.21) can hold only if
ˆ p

(x −x

(t
f
)) = 0 for all x ∈ T
f
,
which is the same as (7.13). Finally (7.12) always implies (7.14), because by the deﬁnition of X(t
f
)
and Exercise 1, ¦Φ(t
f
, t
0
)(z − x

(t
0
)) + x

(t
f
)¦ ∈ X(t
f
) for all z ∈ T
0
, so that from (7.12) we
get
0 ≥ (ˆ p
0
c + ˆ p)

Φ(t
f
, t
0
)(z −x

(t
0
)) for all z ∈ T
0
,
which can hold only if (7.14) holds.
(ii) Now suppose that ˆ p
0
> 0 and ˆ p are such that (7.12), (7.13) are satisﬁed. Let ˜ x ∈ X(t
f
) ∩ T
f
.
Then from (7.13) we conclude that
ˆ p

x

(t
f
) = ˆ p

˜ x ,
so that from (7.12) we get
ˆ p
0
c

x

(t
f
) ≥ ˆ p
0
c

˜ x ;
but then by (7.15) u

() is optimal. ♦
Remark 1: If it is possible to choose ˆ p
0
> 0 then ˆ p
0
= 1, ˆ p = (ˆ p/ˆ p
0
) will also satisfy (7.12),
(7.13), and (7.14). In particular, in part (ii) of the Lemma we may assume ˆ p
0
= 1.
Remark 2: it would be natural to conjecture that in part (i) ˆ p
0
may be chosen > 0. But in Figure
7.3 below, we illustrate a 2-dimensional situation where T
0
= ¦x
0
¦, T
f
is the vertical line, and
T
f
∩ X(t
f
) consists of just one vector. It follows that the control u

() ∈ | for which
x

(t
f
) = φ(t
f
, t
0
, x
0
, u

) ∈ T
f
is optimal for any c. Clearly then for some c (in particular for the
c in Figure 7.3) we are forced to set ˆ p
0
= 0. In higher dimensions the reasons may be more
complicated, but basically if T
f
is “tangent” to X(t
f
) we may be forced to set ˆ p
0
= 0 (see
Exercise 2 below). Finally, we note that part (i) is not too useful if ˆ p
0
= 0, since then (7.12), (7.13),
and (7.14) hold for any vector c whatsoever. Intuitively ˆ p
0
= 0 means that it is so difﬁcult to satisfy
the initial and ﬁnal boundary conditions in (7.11) that optimization becomes a secondary matter.
Remark 3: In (i) the convexity of Ω is only used to guarantee that K(t
f
, t
0
, 0) is convex. But it is
known that K(t
f
, t
0
, 0) is convex even if Ω is not (see Neustadt [1963]).
Exercise 2: Suppose there exists z in the interior of X(t
f
) such that z ∈ T
f
. Then in part (i) we
must have ˆ p
0
> 0.
We now translate the conditions obtained in Lemma 1 in terms of the control u

.
7.2. MORE GENERAL BOUNDARY CONDITIONS 91
Theorem 1: Let x

(t
0
) ∈ T
0
and u

() ∈ |. Let x

(t) = φ(t, t
0
, x

(t
0
), u

) and suppose that
x

(t
f
) ∈ T
f
.
(i) Suppose that Ω is convex. If u

() is optimal for (7.11), then there exist a number p

0
≥ 0, and a
function p

: [t
0
, t
f
] → R
n
, not both identically zero, satisfying

(t) = −A

(t)p

(t) , t
0
≤ t ≤ t
f
(7.22)
initial condition: p

(t
0
)⊥T
0
(x

(t
0
)) (7.23)
ﬁnal condition: (p

(t
f
) −p

0
c)⊥T
f
(x

(t
f
)) , (7.24)
and the maximum principle
H(t, x

(t), u

(t), p

(t)) = M(t, x

(t), p

(t)) , (7.25)
holds for all t ∈ [t
0
, t
f
] except possibly for a ﬁnite set.
(ii) Conversely suppose there exist p

0
> 0 and p

() satisfying (7.22), (7.23), (7.24), and (7.25).
Then u

() is optimal.
[Here
H(t, x, u, p) = p

(A(t)x +B(t)u), M(t, x, p) = sup¦H(t, x, v, p)[v ∈ Ω¦.]
Proof: A repetition of a part of the argument in the proof of Theorem 1 of Section 1 show that if p

satisﬁes (7.22), then (7.25) is equivalent to (7.26):
(p

(t
f
))

x

(t
f
) ≥ (p

(t
f
))

x for all x ∈ K(t
f
, t
0
, x

(t
0
)) . (7.26)
(i) Suppose u

() is optimal and Ω is convex. Then by Lemma 1 there exist ˆ p ≥ 0, ˆ p ∈ R
n
, not
both zero, such that (7.12), (7.13) and (7.14) are satisﬁed. Let p

0
= ˆ p
0
and let p

() be the solution
of (7.22) with the ﬁnal condition
p

(t
f
) = p

0
c + ˆ p = ˆ p
0
c + ˆ p .
Then (7.14) and (7.13) are respectively equivalent to (7.23) and (7.24), whereas since K(t
f
, t
0
, x

(t
0
)) ⊂
X(t
f
), (7.26) is implied by (7.12).
(ii) Suppose p

0
> 0 and (7.22), (7.23), (7.24), and (7.26) are satisﬁed. Let ˆ p
0
= p

0
and ˆ p =
p

(t
f
) −p

0
c, so that (7.24) becomes equivalent to (7.13). Next if x ∈ X(t
f
) we have
(ˆ p
0
c + ˆ p)

x = (p

(t
f
))

x
= (p

(t
f
))

(Φ(t
f
, t
0
)z +w) ,
92 CHAPTER 7. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL
.
x
0
=
T
0
T
f
x

(
t
f
)
=
X
(
t
f
)
¸
T
f
K
(
t
f
,
t
0
,
x
0
)
=
X
(
t
f
)
c
t
Figure 7.3: Situation where ˆ p
0
= 0
7.2. MORE GENERAL BOUNDARY CONDITIONS 93
for some z ∈ T
0
and some w ∈ K(t
f
, t
0
, 0). Hence
(ˆ p
0
c + ˆ p)

x = (p

(
f
))

Φ(t
f
, t
0
)(z −x

(t
0
))
+(p

(t
f
))

(w +φ(t
f
, t
0
)x

(t
0
))
= (p

(t
0
))

(z −x

(t
0
))
+(p

(t
f
))

(w + Φ(t
f
, t
0
)x

(t
0
)) .
But by (7.23) the ﬁrst term on the right vanishes, and since (w+φ(t
f
, t
0
)x

(t
0
)) ∈ K(t
f
, t
0
, x

(t
0
)),
it follows from (7.26) that the second term is bounded by (p

(t
f
))

x

(t
f
). Thus
(ˆ p
0
c + ˆ p)

x

(t
f
) ≥ (ˆ p
0
c + ˆ p)

x for all x ∈ X(t
f
) ,
and so u

() is optimal by Lemma 1. ♦
Exercise 3: Suppose that the control constraint set is Ω(t) which varies continuously with t, and
we require that u(t) ∈ Ω(t) for all t. Show that Theorem 1 also holds for this case where, in (7.25),
M(t, x, p) =sup¦H(t, x, v, p)[v ∈ Ω(t)¦.
Exercise 4: How would you use Exercise 3 to solve Example 3 of Chapter 1?
94 CHAPTER 7. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL
Chapter 8
SEQUENTIAL DECISION PROBLEMS:
CONTINUOUS-TIME OPTIMAL
CONTROL OF NONLINEAR SYSTEMS
We now present a sweeping generalization of the problem studied in the last chapter. Unfortunately
we are forced to omit the proofs of the results since they require a level of mathematical sophis-
tication beyond the scope of these Notes. However, it is possible to convey the main ideas of the
proofs at an intuitive level and we shall do so. (For complete proofs see (Lee and Markus [1967]
or Pontryagin, et al., [1962].) The principal result, which is a direct generalization of Theorem 1 of
7.2 is presented in Section 1. An alternative form of the objective function is discussed in Section
2. Section 3 deals with the minimum-time problem and Section 4 considers the important special
case of linear systems with quadratic cost. Finally, in Section 5 we discuss the so-called singular
case and also analyze Example 4 of Chapter 1.
8.1 Main Results
8.1.1 Preliminary results based on differential equation theory.
We are interested in the optimal control of a system whose dynamics are governed by the nonlinear
differential equation
˙ x(t) = f(t, x, (t), u(t)) , t
0
≤ t ≤ t
f
, (8.1)
where x(t) ∈ R
n
is the state and u(t) ∈ R
p
is the control. Suppose u

() is an optimal control
and x

() is the corresponding trajectory. In the case of linear systems we obtained the necessary
conditions for optimality by comparing x

() with trajectories x() corresponding to other admis-
sible controls u(). This comparison was possible because we had an explicitly characterization of
x() in terms of u(). Unfortunately when f is nonlinear such a characterization is not available.
Instead we shall settle for a comparison between the trajectory x

() and trajectories x() obtained
by perturbing the control u

() and the initial condition x

(t
0
). We can then estimate the difference
between x() and x

() by the solution to a linear differential equation as shown in Lemma 1 below.
But ﬁrst we need to impose some regularity conditions on the differential equation (8.1). We assume
throughout that the function f : [t
0
, t
f
] R
n
R
p
→ R
n
satisﬁes the following conditions:
95
96 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
1. for each ﬁxed t ∈ [t
0
, t
f
], f(t, , ) : R
n
xR
p
→ R
n
is continuously differentiable in the
remaining variables (x, u),
2. except for a ﬁnite subset D ⊂ [t
0
, t
f
], the functions f, f
x
, f
u
are continuous on [t
0
, t
f
] R
n

R
p
, and
3. for every ﬁnite α, there exist ﬁnite number β and γ such that
[f(t, x, u)[ ≤ β +γ[x[ for all t ∈ [t
0
, t
f
], x ∈ R
n
, u ∈ R
p
with [u[ ≤ α .
The following result is proved in every standard treatise on differential equations.
Theorem 1: For every z ∈ R
n
, for every t
1
∈ [t
0
, t
f
], and every piecewise continuous function
u() : [t
0
, t
f
] → R
p
, there exists a unique solution
x(t) = φ(t, t
1
, z, u()) , t
1
≤ t ≤ t
f
,
of the differential equation
˙ x(t) = f(t, x(t), u(t)) , t
1
≤ t ≤ t
f
,
satisfying the initial condition
x(t
1
) = z .
Furthermore, for ﬁxed t
1
≤ t
2
in [t
0
, t
f
] and ﬁxed u(), the function φ(t
2
, t
1
, , u()) : R
n
→ R
n
is
differentiable. Moreover, the n n matrix-valued function Φ deﬁned by
Φ(t
2
, t
1
, z, u()) =
∂φ
∂z
(t
2
, t
1
, z, u())
is the solution of the linear homogeneous differential equation
∂Φ
∂t
(t, t
1
, z, u, ()) = [
∂f
∂x
(t, x, (t), u(t))]Φ(t, t
1
, z, u()), t
1
≤ t ≤ t
f
,
and the initial condition
Φ(t
1
, t
1
, z, u()) = I
n
.
Now let Ω ⊂ R
p
be a ﬁxed set and let | be set of all piecewise continuous functions u() :
[t
0
, t
f
] → Ω. Let u

() ∈ | be ﬁxed and let D

be the set of discontinuity points of u

(). Let
x

0
∈ R
n
be a ﬁxed initial condition.
Deﬁnition: π = (t
1
, . . . , t
m
;
1
, . . . ,
m
; u
1
, . . . , u
m
) is said to be a perturbation data for u

() if
1. m is a nonnegative integer,
2. t
0
< t
1
< t
2
< . . . t
m
< t
f
, and t
i
∈ D

¸
D, i = 1, . . . , m (recall that D is the set of
discontinuity points of f),
3.
i
≥ 0, i = 1, . . . , m, and
4. u
i
∈ Ω, i = 1, . . . , m.
8.1. MAIN RESULTS 97
Let ε(π) > 0 be such that for 0 ≤ ε ≤ ε(π) we have [t
i
− ε
i
, t
i
] ⊂ [t
0
, t
f
] for all i, and
[t
i
−ε
i
, t
i
]
¸
[t
j
−ε
j
, t
j
] = φ for i = j. Then for 0 ≤ ε ≤ ε(π),the perturbed control u
(π,ε)
() ∈ |
corresponding to π is deﬁned by
u
(π,ε)
(t) =

u
i
for all t ∈ [t
i
−ε
i
, t
i
] , i = 1, . . . , m
u

(t) otherwise .
Deﬁnition: Any vector ξ ∈ R
n
is said to be a perturbation for x

0
, and a function x
(ξ,ε)
deﬁned for
ε > 0 is said to be a perturbed initial condition if
lim
ε→0
x
(ξ,ε)
= x

0
,
and
lim
ε→0
1
ε
(x
(ξ,ε)
−x

0
) = ξ .
Now let x

(t) = φ(t, t
0
, x

0
, u

()) and let x
ε
(t) = φ(t, t
0
, x
(ξ,ε)
, u
(π,ε)
()). Let Φ(t
2
, t
1
) =
Φ(t
2
, t
1
, x

(t
1
), u

()). The following lemma gives an estimate of x

(t) − x
ε
(t). The proof of the
lemma is a straightforward exercise in estimating differences of solutions to differential equations,
and it is omitted (see for example (Lee and Markus [1967])).
Lemma 1: lim
ε→0
[x
ε
(t) −x

(t) −εh
(π,ε)
(t)[ = 0 for t ∈ [t
0
, t
1
], where h
(π,ε)
() is given by
h
(π,ε)
(t) = Φ(t, t
0
)ξ , t ∈ [t
0
, t
1
)
= Φ(t, t
0
)ξ + Φ(t, t
1
)[f(t
1
, x

(t
1
), u
1
) −f(t
1
, x

(t
1
), u

(t
1
))]
1
, t ∈ [t
1
, t
2
)
= Φ(t, t
0
)ξ +
i
¸
j=1
Φ(t, t
j
)[f(t
j
, x

(t
j
), u
j
) −f(t
j
, x

(t
j
), u

(t
j
))]
j
, t ∈ [t
i
, t
i+1
)
= Φ(t, t
0
)ξ +
m
¸
j=1
Φ(t, t
m
)[f(t
j
, x

(t
j
), u
j
) −f(t
j
, x

(t
j
), u

t
j
))]
j
, t ∈ [t
m
, t
f
] .
(See Figure 8.1.)
We call h
(π,ξ)
() the linearized (trajectory) perturbation corresponding to (π, ξ).
Deﬁnition: For z ∈ R
n
, t ∈ [t
0
, t
f
] let
K(t, t
0
, z) = ¦φ(t, t
0
, z, u())[u() ∈ |¦
be the set of states reachable at time t, starting at time t
0
in state z, and using controls u() ∈ |.
Deﬁnition: For each t ∈ [t
0
, t
f
], let
Q(t) = ¦h
(π,0)
(t)[πis a perturbation data for u

(), and
h
(π,0)
()is the linearized perturbation
corresponding to(π, 0)¦ .
Remark: By Lemma 1 (x

(t)+εh
(π,ξ)
) belongs to the set K(t, t
0
, x
(ξ,ε)
) up to an error of order o(ε).
In particular for ξ = 0, the set x

(t) + Q(t) can serve as an approximation to the set K(t, t
0
, x

0
).
More precisely we have the following result which we leave as an exercise.
98 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
| | | | | | |
|
|
| |
u
u
1
u
(πε)
()
u
2
u

()
u
3
ε
3
ε
2
ε
1
t
0
t
1
t
2
t
3
t
f
x
x
(
ξ, ε)
x
ε
()
x

()
εh
πξ
t
1
t
2
t
3
t
f
Figure 8.1: Illustration for Lemma 1.
Exercise 1: (Recall the deﬁnition of the tangent cone in 5.1.1.) Show that
Q(t) ⊂ C(K(t, t
0
, x

0
), x

(t)) . (8.2)
We can now prove a generalization of Theorem 1 of 7.1.
Theorem 2: Consider the optimal control problem (8.3):
Maximize ψ(x(t
f
))
subject to
dynamics: ˙ x(t) = f(t, x(t), u(t)) , t
0
≤ t ≤ t
f
,
initial condition: x(t
0
) = x

0
,
ﬁnal condition: x(t
f
) ∈ R
n
,
control constraint: u() ∈ |, i.e., u : [t
0
, t
f
] → Ω and
u() piecewise continuous ,
(8.3)
where ψ : R
n
→ R is differentiable and f satisﬁes the conditions listed earlier.
Let u

() ∈ | be an optimal control and let x

(t) = φ(t, t
0
, x

0
, u

()), t
0
≤ t ≤ t
f
, be the
corresponding trajectory. Let p

(t), t
0
≤ t ≤ t
f
, be the solution of (8.4) and (8.5):

(t) = −[
∂f
∂x
(t, x

(t), u

(t))]

p

(t), t
0
≤ t ≤ t
f
, (8.4)
8.1. MAIN RESULTS 99
ﬁnal condition: p

(t
f
) = ψ(x

(t
f
)) . (8.5)
Then u

() satisﬁes the maximum principle
H(t, x

(t), u

(t), p

(t)) = M(t, x

(t), p

(t)) (8.6)
for all t ∈ [t
0
, t
f
] except possibly for a ﬁnite set. [Here H(t, x, u, p) = p

f(t, x, u, ), M(t, x, p) =
sup¦H(t, x, v, p)[v ∈ Ω¦].
Proof: Since u

() is optimal we must have
ψ(x

(t
f
)) ≥ ψ(z) for all z ∈ K(t
f
, t
0
, x

0
) ,
and so by Lemma 1 of 5.1.1
ψ(x

(t
f
))h ≤ 0 for all h ∈ C(K(t
f
, t
0
, x

0
), x

(t
f
)) ,
and in particular from (8.2)
ψ
x
(x

(t
f
))h ≤ 0 for all h ∈ Q(t
f
) . (8.7)
Now suppose that (8.6) does not hold from some t

∈ D

∪ D. Then there exists v ∈ Ω such that
p

(t

)

[f(t

, x(t

), v) −f(t

, x(t

), u

(t

))] > 0 . (8.8)
If we consider the perturbation data π = (t

; 1; v), then (8.8) is equivalent to
p

(t

)

h
(π,0)
(t

) > 0 . (8.9)
Now from (8.4) we can see that p

(t

)

= p

(t
f
)

Φ(t
f
, t

). Also h
(π,0)
(t
f
) = Φ(t
f
, t

)h
(π,0)
(t

)
so that (8.9) is equivalent to
p

(t
f
)

h
(π,0)
(t
f
) > 0
8.1.2 More general boundary conditions.
In Theorem 2 the initial condition is ﬁxed and the ﬁnal condition is free. The problem involving
more general boundary conditions is much more complicated and requires more reﬁned analysis.
Speciﬁcally, Lemma 1 needs to be extended to Lemma 2 below. But ﬁrst we need some simple
properties of the sets Q(t) which we leave as exercises.
Exercise 2: Show that
(i) Q(t) is a cone, i.e., if h ∈ Q(t) and λ ≥ 0, then λh ∈ Q(t),
(ii) for t
0
≤ t
1
≤ t
2
≤ t
f
, Φ(t
2
, t
1
)Q(t
1
) ⊂ Q(t
2
) .
Deﬁnition: Let C(t) denote the closure of Q(t).
Exercise 3: Show that
(i) C(t) is a convex cone,
(ii) for t
0
≤ t
1
≤ t
2
≤ t
f
, Φ(t
2
, t
1
)C(t
1
) ⊂ C(t
2
) .
Remark: From Lemma 1 we know that if h ∈ C(t) then (x

(t) +εh) belongs to K(t, t
0
, x

(t
0
)) up
to an error of order o(ε). Lemma 2, below, asserts further that if h is in the interior of C(t) then in
fact (x

(t) +εh) ∈ K(t, t
0
, x

(t
0
)) for ε > 0 sufﬁciently small. The proof of the lemma depends
upon some deep topological results and is omitted. Instead we offer a plausibility argument.
Lemma 2: Let h belong to the interior of the cone C(t). Then for all ε > 0 sufﬁciently small,
100 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
(x

(t) +εh) ∈ K(t, t
0
, x

0
) . (8.10)
Plausibility argument. (8.10) is equivalent to
εh ∈ K(t, t
0
, x

(t
0
)) −¦x

(t)¦ , (8.11)
where we have moved the origin to x

(t). The situation is depicted in Figure 8.2.
0
ˆ
C(ε)
ˆ
K(ε)
o(ε)
K(t
1
, t
0
, x

) −¦x

(t)¦
h
C(t)
δε
εh
Figure 8.2: Illustration for Lemma 2.
Let
ˆ
C(ε) be the cross-section of C(t) by a plane orthogonal to h and passing through εh. Let
ˆ
K(ε) be the cross-section of K(t, t
0
, x

0
) −¦x

(t
0
)¦ by the same plane. We note the following:
(i) by Lemma 1 the distance between
ˆ
C(ε) and
ˆ
K(ε) is of the order o(ε);
(ii) since h is in the interior of C(t), the minimum distance between εh and
ˆ
C(ε) is δε where
δ > 0 is independent of ε.
Hence for ε > 0 sufﬁciently small εh must be trapped inside the set
ˆ
K(ε).
(This would constitute a proof except that for the argument to work we need to show that there
are no “holes” in
ˆ
K(ε) through which εh can “escape.” The complications in a rigorous proof arise
precisely from this drawback in our plausibility argument.) ♦
Lemmas 1 and 2 give us a characterization of K(t, t
0
, x

0
) in a neighborhood of x

(t) when we
perturb the control u

() leaving the initial condition ﬁxed. Lemma 3 extends Lemma 2 to the case
when we also allow the initial condition to vary over a ﬁxed surface in a neighborhood of x

0
.
Let g
0
: R
n
→ R

0
be a differentiable function such that the
0
n matrix g
0
x
(x) has rank

0
for all x. Let b
0
∈ R
n
be ﬁxed and let T
0
= ¦x[g
0
(x) − b
0
¦. Suppose that x

0
∈ T
0
and let
T
0
(x

0
) = ¦ξ[g
0
x
(x

0
)ξ = 0¦. Thus, T
0
(x

0
) + ¦x

0
¦ is the plane through x

0
tangent to the surface
T
0
. The proof of Lemma 3 is similar to that of Lemma 2 and is omitted also.
Lemma 3: Let h belong to the interior of the cone ¦C(t)+Φ(t, t
0
)T
0
(x

0
)¦. For ε ≥ 0 let h(ε) ∈ R
n
be such that lim h(ε) = 0, and lim
ε→0
(
1
ε
)h(ε) = h. Then for ε > 0 sufﬁciently small there exists
x
0
(ε) ∈ T
0
such that
(x

(t) +h(ε)) ∈ K(t, t
0
, x
0
(ε)) .
8.1. MAIN RESULTS 101
We can now prove the main result of this chapter. We keep all the notation introduced above.
Further, let g
f
: R
n
→ R

f
be a differentiable function such that g
f
x
(x) has rank
f
for all x.
Let b
f
∈ R
n
be ﬁxed and let T
f
= ¦x[g
f
(x) − b
f
¦. Finally, if x

(t
f
) ∈ T
f
let T
f
(x

(t
f
)) =
¦ξ[g
f
x
(x

(t
f
))ξ = 0¦.
Theorem 3: Consider the optimal control problem (8.12):
Maximize ψ(x(t
f
))
subject to
dynamics: ˙ x(t) = f(t, x(t), u(t)) , t
0
≤ t ≤ t
f
,
initial conditions: g
0
(x(t
0
)) = b
0
,
ﬁnal conditions: g
f
(x(t
f
)) = b
f
,
control constraint: u() ∈ |, i.e., u : [t
0
, t
f
] → Ω and
u() piecewise continuous .
(8.12)
Let u

() ∈ |, let x

0
∈ T
0
and let x

(t) = φ(t, t
0
, x

0
, u

()) be the corresponding trajectory.
Suppose that x

(t
f
) ∈ T
f
, and suppose that (u

(), x

0
) is optimal. Then there exist a number
p

0
≥ 0, and a function p

: [t
0
, t
f
] → R
n
, not both identically zero, satisfying

(t) = −[
∂f
∂x
(t, x

(t), u

(t))]

p

(t), t
0
≤ t ≤ t
f
, (8.13)
initial condition: p

(t
0
)⊥T
0
(x

0
) , (8.14)
ﬁnal condition: (p

(t
f
) −p

0
∇ψ(x

(t
f
)))⊥T
f
(x

(t
f
)) . (8.15)
Furthermore, the maximum principle
H(t, x

(t), u

(t), p

(t)) = M(t, x

(t), p

(t)) (8.16)
holds for all t ∈ [t
0
, t
f
] except possibly for a ﬁnite set. [Here H(t, x, p, u) = p

f(t, x, u, ), M(t, x, p) =
sup¦H(t, x, v, p)[v ∈ Ω¦].
Proof: We break the proof up into a series of steps.
Step 1. By repeating the argument presented in the proof of Theorem 2 we can see that (8.15) is
equivalent to
p

(t
f
)

h ≤ 0 for all h ∈ C(t
f
) . (8.17)
Step 2. Deﬁne two convex sets S
1
, S
2
in R
1+m
as follows:
S
1
= ¦(y, h)[y > 0, h ∈ T
f
(x

(t
f
))¦,
S
2
= ¦(y, h)[y = ψ
x
(x

(t
f
))h, h ∈ ¦C(t
f
) + Φ(t
f
, t
0
)T
0
(x

0
)¦¦ .
We claim that the optimality of (u

(), x

0
) implies that S
1
∩ Relative Interior (S
2
) = φ. Suppose
this is not the case. Then there exists h ∈ T
f
(x

(t
f
)) such that
ψ
x
(x

(t
f
))h > 0 , (8.18)
102 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
h ∈ Interior¦C(t
f
) + Φ(t
f
, t
0
)T
0
(x

0
)¦ . (8.19)
Now by assumption g
f
x
(x

(t
f
) has maximum rank. Since g
f
x
(x

(t
f
))h = 0 it follows that the
Implicit Function Theorem that for ε > 0 sufﬁciently small there exists h(ε) ∈ R
n
such that
g
f
(x

(t
f
) +h(ε)) = b
f
, (8.20)
and, moreover, h(ε) → 0, (1/ε)h(ε) → h as ε → 0. From (8.18) and Lemma 3 it follows that for
ε > 0 sufﬁciently small there exists x
0
(ε) ∈ T
0
and u
ε
() ∈ | such that
x

(t
f
) +h(ε) = φ(t
f
, t
0
, x
0
(ε), u
ε
()) .
Hence we can conclude from (8.20) that the pair (x
0
(ε), u
ε
()) satisﬁes the initial and ﬁnal condi-
tions, and the corresponding value of the objective function is
ψ(x

(t
f
) +h(ε)) = ψ(x

(t
f
)) +ψ
x
(x

(t
f
))h(ε) +o([h(ε)[) ,
and since h(ε) = εh +o(ε) we get
ψ(x

(t
f
) +h(ε)) = ψ(x

(t
f
)) +ε)ψ
x
(x

(t
f
))h +o(ε) ;
but then from (8.18)
ψ(x

(t
f
) +h(ε)) > ψ(x

(t
f
))
for ε > 0 sufﬁciently small, thereby contradicting the optimality of (u

(), x

0
).
Step 3. By the separation theorem for convex sets there exist ˆ p
0
∈ R, ˆ p
1
∈ R
n
, not both zero, such
that
ˆ p
0
y
1
+ ˆ p

1
h
1
≥ ˆ p
0
y
2
+ ˆ p

1
h
2
for all (y
i
, h
i
) ∈ S
1
, i = 1, 2 . (8.21)
Arguing in exactly the same fashion as in the proof of Lemma 1 of 7.2 we can conclude that (8.21)
is equivalent to the following conditions:
ˆ p
0
≥ 0 ,
ˆ p
1
⊥T
f
(x

(t
f
)) ,
(8.22)
Φ(t
f
, t
0
)

(ˆ p
0
∇ψ(x

(t
f
)) + ˆ p
1
)⊥T
0
(x

0
) , (8.23)
and
(ˆ p
0
ψ
x
(x

(t
f
)) + ˆ p

1
)h ≤ 0 for all h ∈ C(t
f
) . (8.24)
If we let ˆ p

0
= ˆ p
0
and p

(t
f
) = ˆ p
0
∇ψ(x

(t
f
)) + ˆ p
1
then (8.22), (8.23), and (8.24) translate respec-
tively into (8.15), (8.14), and (8.17). ♦
8.2. INTEGRAL OBJECTIVE FUNCTION 103
8.2 Integral Objective Function
In many control problems the objective function is not given as a function ψ(x(t
f
)) of the ﬁnal
state, but rather as an integral of the form

t
f
t
0
f
0
(t, x(t), u(t))dt . (8.25)
The dynamics of the state, the boundary conditions, and control constraints are the same as before.
We proceed to show how such objective functions can be treated as a special case of the problems
of the last section. To this end we deﬁned the augmented system with state variable ˜ x = (x
0
, x) ∈
R
1+m
as follows:
·
˜ x=
¸
˙ x
0
(t)
˙ x(t)

=
˜
f(t, ˜ x(t), u(t)) =
¸
f
0
(t, x(t), u(t))
f(t, x(t), u(t))

.
The initial and ﬁnal conditions which are of the form
g
0
(x) = b
0
, g
f
(x) = b
f
are augmented ˜ g
0
(˜ x) =
¸
x
0
g
0
(x)

=
˜
b
0
=
¸
0
b
0

and ˜ g
f
(˜ x) = g
f
(x) = b
f
. Evidently then the problem of maximizing (8.25) is equivalent to the
problem of maximizing
ψ(˜ x(t
f
)) = x
0
(t
f
) ,
subject to the augmented dynamics and constraints which is of the form treated in Theorem 3 of
Section 1, and we get the following result.
Theorem 1: Consider the optimal control problem (8.26):
Maximize

t
f
t
0
f
0
(t, x(t), u(t))dt
subject to
dynamics: ˙ x(t) = f(t, x(t), u(t)), t
0
≤ t ≤ t
f
,
initial conditions: g
0
(x(t
0
)) = b
0
,
ﬁnal conditions: g
f
(x(t
f
)) = b
f
,
control constraint: u() ∈ | .
(8.26)
Let u

() ∈ |, let x

0
∈ T
o
and let x

(t) = φ(t, t
0
, x

0
, u

()), and suppose that x

(t
f
) ∈ T
f
. If
(u

(), x

0
) is optimal, then there exists a function ˜ p

= (p

0
, p

) : [t
0
, t
f
] → R
1+m
, not identically
zero, and with p

0
(t) ≡ constant and p

0
(t) ≥ 0, satisfying
·
˜ p

(t) = −[

˜
f
∂˜ x
(t, x

(t), u

(t))]

˜ p

(t) ,
initial condition: p

(t
0
)⊥T
0
(x

0
) ,
ﬁnal condition: p

(t
f
)⊥T
f
(x

(t
f
)) .
Futhermore, the maximum principle
˜
H(t, x

(t), ˜ p

(t), u

(t)) =
˜
M(t, x

(t), ˜ p

(t))
holds for all t ∈ [t
0
, t
f
] except possibly for a ﬁnite set. [Here
˜
H(t, x, ˜ p, u) = ˜ p

˜
f(t, x, u) =
p
0
f
0
(t, x, u) +p

f(t, x, u), and
˜
M(t, x, ˜ p) = sup¦
˜
H(t, x, ˜ p, v)[v ∈ Ω¦.]
Finally, if f
0
and f do not explicitly depend on t, then
˜
M(t, x

(t), ˜ p

(t)) ≡ constant.
Exercise 1: Prove Theorem 1. (Hint: For the ﬁnal part show that (d/dt)
˜
M(t, x

(t), ˜ p

(t)) ≡ 0.)
104 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
8.3 Variable Final Time
8.3.1 Main result.
In the problem considered up to now the ﬁnal time t
f
is assumed to be ﬁxed. In many important
cases the ﬁnal time is itself a decision variable. One such case is the minimum-time problem where
we want to transfer the state of the system from a given initial state to a speciﬁed ﬁnal state in
minimum time. More generally, consider the optimal control problem (8.27).
Maximize

t
f
t
0
f
0
(t, x(t), u(t))dt
subject to
dynamics: ˙ x(t) = f(t, x, (t), u(t)), , t
0
≤ t ≤ t
f
,
initial condition: g
0
(x(t
0
)) = b
0
,
ﬁnal condition: g
f
(x(t)f)) = b
f
,
control constraint: u() ∈ | ,
ﬁnal-time constraint: t
f
∈ (t
0
, ∞) .
(8.27)
We analyze (8.27) by converting the variable time interval [t
0
, t
f
] into a ﬁxed-time interval [0, 1].
This change of time-scale is achieved by regarding t as a new state variable and selecting a new
time variable s which ranges over [0, 1]. The equation for t is
dt(s)
ds
= α(s) , 0 ≤ s ≤ 1 ,
with initial condition
t(0) = t
0
.
Here α(s) is a new control variable constrained by α(s) ∈ (0, ∞). Now if x() is the solution of
˙ x(t) = f(t, x(t), u(t)) , t
0
≤ t ≤ t
f
, x(t
0
) = x
0
(8.28)
and if we deﬁne
z(s) = x(t(s)), v(s) = u(t(s)) , 0 ≤ s ≤ 1 ,
then it is easy to see that z() is the solution of
dz
ds
(s) = α(s)f(s, z(s), v(s)) , 0 ≤ s ≤ 1 z(0) = x
0
. (8.29)
Conversely from the solution z() of (8.29) we can obtain the solution x() of (8.28) by
x(t) = z(s(t)) , t
0
≤ t ≤ t
f
,
where s() : [t
0
, t
f
] → [0, 1] is the functional inverse of s(t); in fact, s() is the solution of the
differential equation ˙ s(t) = 1/α(s(t)), s(t
0
) = 0.
8.3. VARIABLE FINAL TIME 105
With these ideas in mind it is natural to consider the ﬁxed-ﬁnal-time optimal control problem
(8.30), where the state vector (t, z) ∈ R
1+m
, and the control (α, v) ∈ R
1+p
:
Maximize

1
0
f
0
(t(s), z(s), v(s))α(s)ds
subject to
dynamics: ( ˙ z(s),
˙
t(s)) = (f(t(s), z(s), v(s))α(s), α(s)),
initial constraint: g
0
(z(0)) = b
0
, t(0) = t
0
,
ﬁnal constraint: g
f
(z(1)) = b
f
, t(1) ∈ R ,
control constraint: (v(s), α(s)) ∈ Ω (0, ∞)
for 0 ≤ s ≤ 1 and v(), α() piecewise continuous.
(8.30)
The relation between problems (8.27) and (8.30) is established in the following result.
Lemma 1: (i) Let x

0
∈ T
0
, u

() ∈ |, t

f
∈ (t
0
, ∞) and let x

(t) = φ(t, t
0
, x

0
, u

()) be the
corresponding trajectory. Suppose that x

(t

f
) ∈ T
f
, and suppose that (u

(), x

0
, t

f
) is optimal for
(8.27). Deﬁne z

0
, v

(), and α

() by
z

0
= x

0
v

(s) = u

(t
0
+s(t

f
−t
0
))
α

(s) = (t

f
−t
0
)
, 0 ≤ s ≤ 1 ,
, 0 ≤ s ≤ 1 .
Then ((v

(), α

()), z

0
) is optimal for (8.30).
(ii) Let z

0
∈ T
0
, and let (v

(), α

()) be an admissible control for (8.30) such that the correspond-
ing trajectory (t

(), z

()) satisﬁes the ﬁnal conditions of (8.30). Suppose that ((v

(), α

()), z

0
)
is optimal for (8.30). Deﬁne x

0
, u

() ∈ |, and t

f
by
x

0
= z

0
,
u

(t) = v

(s

(t)) , t
0
≤ t ≤ t

f
,
t

f
= t

(1) ,
where s

() is functional inverse of t

(). Then (u

(), z

0
, t

f
) is optimal for (8.27).
Exercise 1: Prove Lemma 1.
Theorem 1: Let u

() ∈ |, let x

0
∈ T
0
, let t

f
∈ (0, ∞), and let
x

(t) = φ(t, t
0
, x

0
, u

()), t
0
≤ t ≤ t
f
, and suppose that x

(t

f
) ∈ T
f
. If (u

(), x

0
, t

f
) is optimal
for (8.27), then there exists a function ˜ p

= (p

0
, p

) : [t
0
, t

f
] → R
1+m
, not identically zero, and
with p

0
(t) ≡ constant and p

0
(t) ≥ 0, satisfying
·
˜ p

(t) = −[

˜
f
∂˜ x
(t, x

(t), u

(t))]

˜ p

(t) ,
(8.31)
initial condition: p

(t
0
)⊥T
0
(x

0
) , (8.32)
106 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
ﬁnal condition: p

(t

f
)⊥T
f
(x

(t

f
)) . (8.33)
Also the maximum principle
˜
H(t, x

(t), ˜ p

(t), u

(t)) =
˜
M(t, x

(t), ˜ p

(t)) , (8.34)
holds for all t ∈ [t
0
, t
f
] except possibly for a ﬁnite set. Furthermore, t

f
must be such that
ˆ
H(t

f
, x

(t

f
), ˜ p

(t

f
), u

(t

f
)) = 0 . (8.35)
Finally, if f
0
and f do not explicitly depend on t, then
ˆ
M(t, x

(t), ˜ p

(t)) ≡ 0.
Proof: By Lemma 1, z

0
= x

0
, v

(s) = u

(t
0
+s(t

f
−t
0
)) and α

(s) = (t

f
−t
0
) for 0 ≤ s ≤ 1
constitute an optimal solution for (8.30). The resulting trajectory is
z

(s) = x

(t
0
+s(t

f
−t
0
)), t

(s) = t
0
+s(t

f
−t
0
), 0 ≤ s ≤ 1 , so that in particular
z

(1) = x

(t

f
).
By Theorem 1 of Section 2, there exists a function
˜
λ

= (λ

0
, λ

, λ

n+1
) : [0, 1] → R
1+n+1
, not
identically zero, and with λ

0
(s) ≡ constant and λ

0
(s) ≥ 0, satisfying

˙
λ

0
(t)
˙
λ

(t)
˙
λ

n+1
(t)
¸
¸
¸
¸
¸
¸
= −

0
¦[
∂f
0
∂z
(t

(s), z

(s), v

(s))]

λ

0
(s)
+[
∂f
∂z
(t

(s), z

(s), v

(s))]

λ

(s)¦α

(s)
¦[
∂f
0
∂t
(t

(s), z

(s), v

(s))]

λ

0
(s)
+[
∂f
∂t
(t

(s), z

(s), v

(s))]

λ

(s)¦α

(s)
¸
¸
¸
¸
¸
¸
¸
(8.36)
initial condition: λ

(0)⊥T
0
(z

0
) (8.37)
ﬁnal condition: λ

(1)⊥T
f
(z

(1)) , λ

n+1
(1) = 0 . (8.38)
Furthermore, the maximum principle
λ

0
(s)f
0
(t

(s), z

(s), v

(s))α

(s)

(s)

f(t

(s), z

(s), v

(s))α

(s) +λ

n+1
(s)α

(s)
= sup¦[λ

0
(s)f
0
(t

(s), z

(s), w)β

(s)

f(t

(s), z

(s), w)β +λ

n+1
(s)β][w ∈ Ω, β ∈ (0, ∞)¦
(8.39)
holds for all s ∈ [0, 1] except possibly for a ﬁnite set.
Let s

(t) = (t −t
0
)/(t

f
−t
0
), t
0
≤ t ≤ t

f
, and deﬁne ˜ p

= (p

0
, p

) : [t
0
, t

f
] → R
1+n
by
p

0
(t) = λ

0
(s

(t)), p

(t) = λ

(s

(t)), t
0
≤ t ≤ t

f
. (8.40)
First of all, ˜ p

is not identically zero. Because if ˜ p

≡ 0, then from (8.40) we have (λ

0
, λ

) ≡ 0 and
then from (8.36), λ

n+1
≡ constant, but from (8.38), λ

n+1
(1) = 0 so that we would have
˜
λ

≡ 0
8.3. VARIABLE FINAL TIME 107
which is a contradiction. It is trivial to verify that ˜ p

() satisﬁes (8.31), and, on the other hand (8.37)
and (8.38) respectively imply (8.32) and (8.33). Next, (8.39) is equivalent to
λ

0
(s)f
0
(t

(s), z

(s), v

(s))

(s)

f(t

(s), z

(s), v

(s)) +λ

n+1
(s) = 0
(8.41)
and
λ

0
(s)f
0
(t

(s), z

(s), v

(s)) +λ

(s)

f(t

(s), z

(s), v

(s))
= Sup ¦[λ

0
(s)f
0
(t

(s), z

(s), w) +λ

(s)

f(t

(s), z

(s), w)][w ∈ Ω¦.
(8.42)
Evidently (8.42) is equivalent to (8.34) and (8.35) follows from (8.41) and the fact that λ

n+1
(1) = 0.
Finally, the last assertion of the Theorem follows from (8.35) and the fact that
˜
M(t, x

(t), ˜ p

(t)) ≡
constant if f
0
, f are not explicitly dependent on t. ♦
8.3.2 Minimum-time problems
.
We consider the following special case of (8.27):
Maximize

t
f
t
0
(−1)dt
subject to
dynamics: ˙ x(t) = f(t, x(t), u(t)), t
0
≤ t ≤ t
f
initial condition: x(t
0
) = x
0
,
ﬁnal condition: x(t
f
) = x
f
,
control constraint: u() ∈ | ,
ﬁnal-time constraint: t
f
∈ (t
0
, ∞) .
(8.43)
In (8.43), x
0
, x
f
are ﬁxed, so that the optimal control problem consists of ﬁnding a control which
transfers the system from state x
0
at time t
0
to state x
f
in minimum time. Applying Theorem 1 to
this problem gives Theorem 2.
Theorem 2: Let t

f
∈ (t
0
, ∞) and let u

: [t
0
, t

f
] → Ω be optimal. Let x

() be the corresponding
trajectory. Then there exists a function p

: [t
0
, t

f
] → R
n
, not identically zero, satisfying

(t) = −[
∂f
∂x
(t, x

(t), u

(t))]

p

(t), t
0
≤ t ≤ t

f
,
initial condition: p

(t
0
) ∈ R
n
,
ﬁnal condition: p

(t

f
) ∈ R
n
.
Also the maximum principle
H(t, x

(t), p

(t), u

(t)) = M(t, x

(t), p

(t)) (8.44)
holds for all t ∈ [t
0
, t

f
] except possibly for a ﬁnite set.
Finally,
M(t

f
, x

(t
f
), p

(t
f
)) ≥ 0 (8.45)
and if f does not depend explicitly on t then
M(t, x

(t), p

, (t)) ≡ constant . (8.46)
108 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
Exercise 2: Prove Theorem 2.
We now study a simple example illustrating Theorem 2. Example 1: The motion of a particle is
described by
m¨ x(t) +σ ˙ x(t) = u(t) ,
where m = mass, σ = coefﬁcient of friction, u = applied force, and x = position of the particle. For
simplicity we suppose that x ∈ R, u ∈ R and u(t) constrained by [u(t)[ ≤ 1. Starting with an
initial condition x(0) = x
01
, ˙ x(0) = x
02
we wish to ﬁnd an admissible control which brings the
particle to the state x = 0, ˙ x = 0 in minimum time.
Solution: Taking x
1
= x, x
2
= ˙ x we rewrite the particle dynamics as
¸
˙ x
1
(t)
˙ x
2
(t)

=
¸
0 1
0 −α
¸
x
1
(t)
x
2
(t)

+
¸
0
b

u(t) , (8.47)
where α = (σ/m) > 0 and b = (1/m) > 0. The control constraint set is Ω = [−1, 1].
Suppose that u

() is optimal and x

() is the corresponding trajectory. By Theorem 2 there exists
a non-zero solution p

() of
¸
˙ p

1
(t)
˙ p

2
(t)

= −
¸
0 0
1 −α
¸
p

1
(t)
p

2
(t)

(8.48)
such that (8.44), (8.45), and (8.46) hold. Now the transition matrix function of the homogeneous
part of (8.47) is
Φ(t, τ) =
¸
1
1
α
(1 −e
−α(t−τ)
)
0 e
−α(t−τ)

,
so that the solution of (8.48) is
¸
p

1
(t)
p

2
(t)

=
¸
1 0
1
α
(1 −e
αt
) e
αt
¸
p

1
(0)
p

2
(0)

,
or
p

1
(t) ≡ p

1
(0) ,
and
p

2
(t) =
1
α
p

1
(0) +e
αt
(−
1
α
p

1
(0) +p

2
(0)) . (8.49)
The Hamiltonian H is given by
H(x

(t), p

(t), v) = (p

1
(t) −αp

2
(t))x

2
(t) +bp

2
(t)v
= e
αt
(p

1
(0) −αp

2
(0))x

2
(t) +pb

2
(t)v ,
8.3. VARIABLE FINAL TIME 109
so that from the maximum principle we can immediately conclude that
u

(t) =

+1 if p

2
(t) > 0,
−1 if p

2
(t) < 0,
? if p

2
(t) = 0 .
(8.50)
Furthermore, since the right-hand side of (8.47) does not depend on t explicitly we must also have
e
αt
(p

1
(0) −αp

2
(0))x

2
(t) +bp

2
(t)u

(t) ≡ constant. (8.51)
We now proceed to analyze the consequences of (8.49) and (8.50). First of all since p

1
(t) ≡
p

1
(0), p

2
() can have three qualitatively different forms.
Case 1. −p

1
(0) + αp

2
(0) > 0: Evidently then, from (8.49) we see that p

2
(t) must be a strictly
monotonically increasing function so that from (8.50) u

() can behave in one of two ways:
either
u

(t) =

−1 for t <
ˆ
t and p

2
(t) < 0 for t <
ˆ
t,
+1 for t >
ˆ
t and p

2
(t) > 0 for t >
ˆ
t,
or
u

(t) ≡ +1 and p

2
(t) > 0 for all t.
Case 2. −p

1
(0) +αp

2
(0) < 0 : Evidently u

() can behave in one of two ways:
either
u

(t) =

+1 for t <
ˆ
t and p

2
(t) > 0 for t <
ˆ
t,
−1 for t >
ˆ
t and p

2
(t) < 0 for t >
ˆ
t,
or
u

(t) ≡ −1 and p

(t) < 0 for all t.
Case 3. −p

1
(0) + αp

2
(0) = 0 : In this case p

2
(t) ≡ (1/α)p

1
(0). Also since p

(t) ≡ 0, we must
have in this case p

1
(0) = 0. Hence u

() we can behave in one of two ways:
either
u

(t) ≡ +1 and p

2
(t) ≡
1
α
p

1
(0) > 0 ,
or
u

(t) ≡ −1 and p

2
(t) ≡
1
α
p

1
(0) < 0 ,
110 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
Thus, the optimal control u

is always equal to +1 or -1 and it can switch at most once between
these two values. The optimal control is given by
u

(t) = sgn p

2
(t)
= sgn [
1
α
p

1
(0) +e
αt
(−
1
α
p

1
(0) +p

2
(0))] .
Thus the search for the optimal control reduces to ﬁnding p

1
(0), p

2
(0) such that the solution of the
differential equation
˙ x = x
2
˙ x
2
= −αx
2
+b sgn[
1
α
p

1
(0) +e
αt
(−
1
α
p

1
(0) +p

2
(0))] ,
(8.52)
with initial condition
x
1
(0) = x
10
, x
20
= x
20
(8.53)
also satisﬁes the ﬁnal condition
x
1
(t

f
) = 0, x
2
(t

f
) = 0 , (8.54)
for some t

f
> 0; and then t

f
is the minimum time.
There are at least two ways of solving the two-point boundary value problem (8.52), (8.53), and
(8.54). One way is to guess at the value of p

(0) and then integrate (8.52) and (8.53) forward in time
and check if (8.54) is satisﬁed. If (8.54) is not satisﬁed then modify p

(0) and repeat. An alternative
is to guess at the value of p

(0) and then integrate (8.52) and (8.54) backward in time and check of
(8.53) is satisﬁed. The latter approach is more advantageous because we know that any trajectory
obtained by this procedure is optimal for initial conditions which lie on the trajectory. Let us follow
this procedure.
Suppose we choose p

(0) such that −p

1
(0) = αp

2
(0) = 0 and p

2
(0) > 0. Then we must have
u

(t) ≡ 1. Integrating (8.52) and (8.54) backward in time give us a trajectory ξ(t) where
˙
ξ
1
(t) = −
˙
ξ
2
(t)
˙
ξ
2
(t) = αξ
2
(t) −b ,
with
ξ
1
(0) −ξ
2
(0) = 0 .
This gives
ξ
1
(t) =
b
α
(−t +
e
αt
−1
α
) , ξ
2
(t) =
b
α
(1 −e
αt
) ,
which is the curve OA in Figure 8.3.
On the other hand, if p

(0) is such that −p

1
(0) + αp

2
(0) = 0 and p

2
(0) < 0, then u

(t) ≡ −1
and we get
ξ
1
(t) = −
b
α
(−t +
e
αt
−1
α
) , ξ
2
(t) = −
b
α
(1 −e
αt
) ,
which is the curve OB.
8.3. VARIABLE FINAL TIME 111
B
u

≡ −1
D
u

≡ 1
C
ξ
1
O
ξ
2
u

≡ 1
A
E
u

≡ −1
F
Figure 8.3: Backward integration of (8.52) and (8.54).
Next suppose p

(0) is such that −p

1
(0) + αp

2
(0) > 0, and p

2
(0) < 0. Then [(1/α)p

1
(0) +
e
αt
(−(1/α)p

1
(0) + p

2
(0))] will have a negative value for t ∈ (0,
ˆ
t) and a positive value for t ∈
(
ˆ
t, ∞). Hence, if we integrate (8.52), (8.54) backwards in time we get trajectory ξ(t) where
˙
ξ(t) = −ξ
2
(t)
˙
ξ
2
(t) = αξ
2
(t)+

−b for t <
ˆ
t
b for t >
ˆ
t ,
with ξ
1
(0) = 0, ξ
2
(0) = 0. This give us the curve OCD. Finally if p

(0) is such that −p

1
(0) +
αp

2
(0) < 0, and p

2
(0) < 0, then u

(t) = 1 for t <
ˆ
t and u

(t) = −1 for t >
ˆ
t, and we get the
curve OEF.
We see then that the optimal control u

() has the following characterizing properties:
u

(t) =

1 if x

(t) is above BOA or on OA
−1 if x

(t) is below BOA or on OB .
Hence we can synthesize the optimal control in feedback from: u

(t) = ψ(x

(t)) where the
B
u

≡ −1
x
2
u

≡ 1
x
1
A
u

≡ 1
O
u

≡ −1
Figure 8.4: Optimal trajectories of Example 1.
112 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
function ψ : R
2
→ ¦1, −1¦ is given by (see Figure 8.4)
ψ(x
1
, x
2
) =

1 if (x
1
, x
2
) is above BOA or on OA
−1 if (x
1
, x
2
) is below BOA or on OB .
An important class of problems which arise in practice is the case when the dynamics are linear and
the objective function is quadratic. Speciﬁcally, consider the optimal control problem (8.55):
Minimize

T
0
1
2
[x

(t)P(t)x(t) +u

(t)Q(t)u(t)]dt
subject to
dynamics: ˙ x(t) = A(t)x(t) +B(t)u(t), 0 ≤ t ≤ T ,
initial condition: x(0) = x
0
,
ﬁnal condition: G
f
x(t) = b
f
,
control constraint: u(t) ∈ R
p
, u() piecewise continuous.
(8.55)
In (8.56) we assume that P(t) is an n n symmetric, positive semi-deﬁnite matrix whereas Q(t) is
a p p symmetric, positive deﬁnite matrix. G
f
is a given
f
n matrix, and x
0
∈ R
n
, b
f
∈ R

f
are given vectors. T is a ﬁxed ﬁnal time.
We apply Theorem 1 of Section 2, so that we must search for a number p

0
≥ 0 and a function
p

: [0, T] → R
n
, not both zero, such that
˙ p

(t) = −p

0
(−P(t)x

(t)) −A

(t)p

(t) , (8.56)
and
p

(t)⊥T
f
(x

(t)) = ¦ξ[G
f
ξ = 0¦ . (8.57)
The Hamiltonian function is
H(t, x

(t), ˜ p

(t), v) = −
1
2
p

0
[x

(t)

P(t)x

(t) +v

Q(t)v]
+p

(t)

[A(t)x

(t) +B(t)v]
so that the optimal control u

(t) must maximize

1
2
p

0
v

Q(t)v +p

(t)

B(t)v for v ∈ R
p
. (8.58)
If p

0
> 0, this will imply
u

(t) =
1
p

0
Q
−1
(t)B

(t)p

(t) ,
(8.59)
whereas if p

0
= 0, then we must have
p

(t)

B(t) ≡ 0 (8.60)
because otherwise (8.58) cannot have a maximum.
8.5. THE SINGULAR CASE 113
We make the following assumption about the system dynamics.
Assumption: The control system ˙ x(t) = A(t)x(t) + B(t)u(t) is controllable over the interval
[0, T]. (See (Desoer [1970]) for a deﬁnition of controllability and for the properties we use below.)
Let Φ(t, τ) be the transition matrix function of the homogeneous linear differential equation ˙ x(t) =
A(t)x(t). Then the controllability assumption is equivalent to the statement that for any ξ ∈ R
n
ξ

Φ(t, τ)B(τ) = 0 , 0 ≤ τ ≤ T , implies ξ = 0 . (8.61)
Next we claim that if the system is controllable then p

0
= 0, because if p

0
= 0 then from (8.56)
we can see that
p

(t) = (Φ(T, t))

p

(T)
and hence from (8.60)
(p

(t))

Φ(T, t)B(t) = 0 , 0 ≤ t ≤ T ,
but then from (8.61) we get p

(T) = 0. Hence if p

0
= 0, then we must have ˜ p

(t) ≡ 0 which is a
contradiction. Thus, under the controllability assumption, p

0
> 0, and hence the optimal control is
given by (8.59). Now if p

0
> 0 it is trivial that ˆ p

(t) = (1, (p

(t)/p

0
)) will satisfy all the necessary
conditions so that we can assume that p

0
= 1. The optimal trajectory and the optimal control is
obtained by solving the following two-point boundary value problem:
˙ x

(t) = A(t)x

(t) +B(t)Q
−1
(t)B

(t)p

(t)
˙ p(t) = P(t)x

(t) −A

(t)p

(t)
x

(0) = x
0
, G
f
x

(T) = b
f
, p

(T)⊥T
f
(x

(T)) .
For further details regarding the solution of this boundary value problem and for related topics see
(See and Markus [1967]).
8.5 The Singular Case
In applying the necessary conditions derived in this chapter it sometimes happens that H(t, x

(t), p

(t), v)
is independent of v for values of t lying in a non-zero interval. In such cases the maximum principle
does not help in selecting the optimal value of the control. We are faced with the so-called singular
case (because we are in trouble–not because the situation is rare). We illustrate this by analyzing
Example 4 of Chapter 1.
The problem can be summarized as follows:
Maximize

T
0
c(t)dt =

T
0
(1 −s(t))f(k(t))dt
subject to
dynamics:
˙
k(t) = s(t)f(k(t)) −µk(t) , 0 ≤ t ≤ T
initial constraint: k(0) = k
0
,
ﬁnal constraint: k(t) ∈ R ,
control constraint: s(t) ∈ [0, 1], s() piecewise continuous.
114 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
We make the following assumptions regarding the production function f:
f
k
(k) > 0, f
kk
(K) < 0 for all k , (8.62)
lim
k→0
f
k
(k) = ∞ .
(8.63)
Assumption (8.62) says that the marginal product of capital is positive and this marginal product
decreases with increasing capital. Assumption (8.63) is mainly for technical convenience and can
be dispensed with without difﬁculty.
Now suppose that s

: [0, T] → [0, 1] is an optimal savings policy and let k

(t), 0 ≤ t ≤ T,
be the corresponding trajectory of the capital-to-labor ratio. Then by Theorem 1 of Section 2, there
exist a number p

0
≥ 0, and a function p

: [0, T] → R, not both identically zero, such that
˙ p

(t) = −p

0
(1 −s

(t))f
k
(k

(t)) −p

(t)[s

(t)f
k
(k

(t)) −µ] (8.64)
with the ﬁnal condition
p

(T) = 0 , (8.65)
and the maximum principle holds. First of all, if p

0
= 0 then from (8.64) and (8.65) we must also
have p

(t) ≡ 0. Hence we must have p

0
> 0 and then by replacing (p

0
, p

) by (1/p

0
)(p

0
, p

) we
can assume without losing generality that p

0
= 1, so that (8.64) simpliﬁes to
˙ p

(t) = −1(1 −s

(t))f
k
(k

(t)) −p

(t)[s

(t)f
k
(k

(t)) −µ] . (8.66)
The maximum principle says that
H(t, k

(t), p

(t), s) = (1 −s)f(k

(t)) +p

(t)[sf(k

(t)) −µk

(t)]
is maximized over s ∈ [0, 1] at s

(t), which immediately implies that
s

(t) =

1 if p

(t) > 1
0 if p

(t) < 1
? if p

(t) = 1
(8.67)
We analyze separately the three cases above.
Case 1. p

(t) > 1, s

(t) = 1 : Then the dynamic equations become
˙
k

(t) = f(k

(t)) −µk

(t) ,
˙ p

(t) = −p

(t)[f
k
(k

(t)) −µ] .
(8.68)
The behavior of the solutions of (8.68) is depicted in the (k, p)−, (k, t)− and (p, t)−planes in
Figure 8.5. Here k
G
, k
H
are the solutions of f
k
(k
G
) −µ = 0 and f(k
M
) −µk = 0. Such solutions
exist and are unique by virtue of the assumptions (8.62) and (8.63). Futhermore, we note from
(8.62) that k
G
< k
M
, and f
k
(k) − µ
<
> 0 according as k
<
> k
G
whereas f(k) − µk
>
< 0 according
as k
<
> k
M
. (See Figure 8.6.)
8.5. THE SINGULAR CASE 115
p
f
k
> µ
f
k
< µ
f < µk f > µk
k
k
M
k
G
l
k
k
M
t
p
l
t
Figure 8.5: Illustration for Case 1.
Case 2. p

(t) < 1, s

(t) = 0: Then the dynamic equations are
˙
k

(t) = −µk

(t) ,
˙ p

(t) = −f
k
(k

(t)) +µp

(t) ,
giving rise to the behavior illustrated in Figure 8.7.
Case 3. p

(t) = 1, s

(t) =?: (Possibly singular case.) Evidently if p

(t) = 1 only for a ﬁnite set of
times t then we do not have to worry about this case. We face the singular case only if p

(t) = 1
for t ∈ I, where I is a non-zero interval. But then we have ˙ p

(t) = 0 for t ∈ I so that from (8.66)
we get
−(1 −s

(t))f
k
(k

(t)) −[s

(t)f
k
(k

(t)) −µ] = 0 for t ∈ I ,
so
−f
k
(k

(t)) +µ = 0 for t ∈ I ,
or
k

(t) = k
G
for t ∈ I . (8.69)
In turn then we must have
˙
k

(t) = 0 for t ∈ I so that
s

(t)f(k
G
) −µK
G
= 0 for t ∈ I ,
and hence,
s

(t) = µ
k
G
f(k
G
)
for t ∈ I . (8.70)
116 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
.
f
line of slope µ
µk
f(k)
k
k
M
k
G
Figure 8.6: Illustration for assumptions (8.62), (8.63).
Thus in the singular case the optimal solution is characterized by (8.69) and (8.70), as in Figure 8.8.
We can now assemble separate cases to obtain the optimal control. First of all, from the ﬁnal
condition (8.65) we know that for t close to T, p

(t) < 1 so that we are in Case 2. We face two
possibilities: Either (A)
p

(t) < 1 for all t < [0, T]
and then s

(t) = 0, k

(t) = k
0
e
−µt
, for 0 ≤ t ≤ T, or (B)
there exists t
2
∈ (0, T) such that p

(t
2
) = 1 and p

(t) < 1 for t
2
< t ≤ T .
We then have three possibilities depending on the value of k

(t
2
):
(Bi) k

(t
2
) < k
G
: then ˙ p

(t
2
) < 0 so that p

(t) > 1 for t < t
2
and we are in Case 1 so that
s

(t) = 1 for t < t
2
. In particular we must have k
0
< k
G
.
(Bii) k

(t
2
) > k
G
: then ˙ p

(
2
) > 0 but then p

(t
2
+ ε) > 1 for ε > 0 sufﬁciently small and since
p

(T) = 0 there must exist t
3
∈ (t
2
, T) such that p

(t
3
) = 1. This contradicts the deﬁnition of t
2
so that this possibility cannot arise.
(Biii) k

(t
2
) − k
G
: then we can have a singular arc in some interval (t
1
, t
2
) so that p

(t) =
1, k

(t) = k
G
, and s

(t) = µ(k
G
/f(k
G
)) for t ∈ (t
1
, t
2
). For t < t
1
we either have p

(t) >
1, s

(t) > 1 if k
0
< k
G
, or we have p

(t) < 1, s

(t) = 0 if k > k
G
.
The various possibilities are illustrated in Figure 8.9.
The capital-to-labor ratio k
G
is called the golden mean and the singular solution is called the
golden path. The reason for this term is contained in the following exercise.
Exercise 1: A capital-to-labor ratio
ˆ
k is said to be sustainable if there exists ˆ s ∈ [0, 1] such that
ˆ sf(
ˆ
k) −µ
ˆ
k = 0. Show that k
G
is the unique sustainable capital-to-labor ratio which maximizes
sustainable consumption (1 −s)f(k).
8.6. BIBLIOGRAPHICAL REMARKS 117
p
k
k
G
l
k
t
p
t
l
Figure 8.7: Illustration for Case 2.
8.6 Bibliographical Remarks
The results presented in this chapter appeared in English in full detail for the ﬁrst time in 1962 in the
book by Pontryagin, et al., cited earlier. That book contains many extensions and many examples
and it is still an important source. However, the derivation of the maximum principle given in the
book by Lee and Markus is more satisfactory. Several important generalizations of the maximum
principle have appeared. On the one hand these include extensions to inﬁnite-dimensional state
spaces and on the other hand they allow for constraints on the state more general than merely initial
and ﬁnal constraints. For a uniﬁed, but mathematically difﬁcult, treatment see (Neustadt [1969]).
For a less rigorous treatment of state-space constraints see (Jacobson, et al, [1971]), whereas for a
discussion of the singular case consult (Kelley, et al. [1968]).
For an applications-oriented treatment of this subject the reader is referred to (Athans and Falb
[1966]) and (Bryson and Ho [1969]). For applications of the maximum principle to optimal eco-
nomic growth see (Shell [1967]). There is no single source of computational methods for optimal
control problems. Among the many useful techniques which have been proposed see (Lasdon, et
al., [1967]), (Kelley [1962]), (McReynolds [1966]), and (Balakrishnan and Neustadt [1964]); also
consult (Jacobson and Mayne [1970]), and (Polak [1971]).
118 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
.
p
k
k
G
1
k
t
k
G
p
t
1
Figure 8.8: Case 3. The singular case.
8.6. BIBLIOGRAPHICAL REMARKS 119
. .
.
.
.
.
. .
p

1
t
T
s

1
t
T
k

t
T
p

Case (A)
t
T t
2
t
1
s

1
µk
G
f(k
G
)
t
k

k
G
k
0
t
p

1
t
T
t
2
s

1
t
T
t
2
k

t
T
t
2
p

Case (Bi)
t
T
t
2
t
1
s

t
k

t
Case (Biii)
Figure 8.9: The optimal solution of example.
120 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
Chapter 9
Dynamic programing
SEQUENTIAL DECISION PROBLEMS: DYNAMIC PROGRAMMING FORMULATION
The sequential decision problems discussed in the last three Chapters were analyzed by varia-
tional methods, i.e., the necessary conditions for optimality were obtained by comparing the op-
timal decision with decisions in a small neighborhood of the optimum. Dynamic programming
(DP is a technique which compares the optimal decision with all the other decisions. This global
comparison, therefore, leads to optimality conditions which are sufﬁcient. The main advantage of
DP, besides the fact that it give sufﬁciency conditions, is that DP permits very general problem for-
mulations which do not require differentiability or convexity conditions or even the restriction to a
ﬁnite-dimensional state space. The only disadvantage (which unfortunately often rules out its use)
of DP is that it can easily give rise to enormous computational requirements.
In the ﬁrst section we develop the main recursion equation of DP for discrete-time problems. The
second section deals with the continuous-time problem. Some general remarks and bibliographical
references are collected in the ﬁnal section.
9.1 Discrete-time DP
We consider a problem formulation similar to that of Chapter VI. However, for notational conve-
nience we neglect ﬁnal conditions and state-space constraints.
Maximize
N−1
¸
i=0
f
0
(i, x(i), u(i)) + Φ(x(N))
subject to
dynamics: x(i + 1) = f(i, x(i), u(i)) , i = 0, 1, . . . , N −1 ,
initial condition: x(0) = x
0
,
control constraint: u(i) ∈ Ω
i
, i = 0, 1, . . . , N −1 .
(9.1)
In (9.1), the state x(i) and the control u(i) belong to arbitrary sets X and U respectively. X and U
may be ﬁnite sets, or ﬁnite-dimensional vector spaces (as in the previous chapters), or even inﬁnite-
dimensional spaces. x
0
∈ X is ﬁxed. The Ω
i
are ﬁxed subsets of U. Finally f
0
(i, , ) : X U →
R, Φ : X → R, f(i, , ) : X U → X are ﬁxed functions.
121
122 CHAPTER 9. DYNAMIC PROGRAMING
The main idea underlying DP involves embedding the optimal control problem (9.1), in which
the system starts in state x
0
at time 0, into a family of optimal control problems with the same
dynamics, objective function, and control constraint as in (9.1) but with different initial states and
initial times. More precisely, for each x ∈ X and k between ) and N − 1, consider the following
problem:
Maximize
N−1
¸
i=k
f
0
(i, x(i), u(i)) + Φ(x(N)) ,
subject to
dynamics: x(i + 1) = f(i, x(i), u(i)), i = k, k + 1, . . . , N −1,
initial condition: x(k) = x,
control constraint: u(i) ∈ Ω
i
, i = k, k + 1, , N −1 .
(9.2)
Since the initial time k and initial state x are the only parameters in the problem above, we will
sometimes use the index (9.2)
k,x
to distinguish between different problems. We begin with an
elementary but crucial observation.
Lemma 1: Suppose u

(k), . . . , u

(N − 1) is an optimal control for (9.2)
k,x
, and let x

(k) =
x, x

(k + 1), . . . , x

(N) be the corresponding optimal trajectory. Then for any , k ≤ ≤ N −
1, u

(), . . . , u

(N −1) is an optimal control for (9.2)
,x

()
.
Proof: Suppose not. Then there exists a control ˆ u(), ˆ u( + 1), . . . , ˆ u(N − 1), with corresponding
trajectory ˆ x() = x

(), ˆ x( + 1), . . . , ˆ x(N), such that
N−1
¸
i=
f
0
(i, ˆ x(i), ˆ u(i)) + Φ(ˆ x(N))
>
N−1
¸
i=
f
0
(i, x

(i), u

(i)) + Φ(x

(N)) .
(9.3)
But then consider the control ˜ u(k), . . . , ˜ u(N −1) with
˜ u(i)

u

(i) , i = k, . . . , −1
ˆ u(i) , i = , . . . , N −1 ,
and the corresponding trajectory, starting in state x at time k, is ˜ x(k), . . . , ˜ x(N) where
˜ x(i) =

x

(i) , i = k, . . . ,
ˆ x(i) , i = + 1, . . . , N .
The value of the objective function corresponding to this control for the problem (9.2)
k,x
is
N−1
¸
i=k
f
0
(i, ˜ x(i), ˜ u(i)) + Φ(˜ x(n))
=
−1
¸
i=k
f
0
(i, x

(i), u

(i)) +
N−1
¸
i=
f
0
(i, ˆ x(i), ˆ u(i)) + Φ(ˆ x(N))
>
N−1
¸
i=k
f
0
(i, x

(i), u

(i)) + Φ(x

(N)) ,
9.1. DISCRETE-TIME DP 123
by (9.3), so that u

(k), . . . , u

(N −1) cannot be optimal for 9.2)
k
(end theorem)
From now on we assume that an optimal solution to (9.2)
k,x
exists for all 0 ≤ k ≤ N −1, and all
x ∈ X. Let V (k, x) be the maximum value of (9.2)
k,x
. We call V the (maximum) value function.
Theorem 1: Deﬁne V (N, ) by (V (N, x) = Φ(x). V (k, x) satisﬁes the backward recursion equa-
tion
V (k, x) = Max¦f
0
, (k, x, u) +V (k
1
, f(k, x, u, ))[u ∈ Ω
k
¦, 0 ≤ k ≤ N −1 . (9.4)
Proof: Let x ∈ X, let u

(k), . . . , u

(N − 1) be an optimal control for (9.2)
k,x
, and let x

(k) =
x, . . . , x

(N) be the corresponding trajectory be x(k) = x, . . . , x(N). We have
N−1
¸
i=k
f
0
(i, x

(i), u

(i)) + Φ(x

(N))

N−1
¸
i=k
f
0
(i, x(i), u(i)) + Φ(x(N)) .
(9.5)
By Lemma 1 the left-hand side of (9.5) is equal to
f
0
(k, x, u

(k)) +V (k + 1, f(k, x

, u

(k)) .
On the other hand, by the deﬁnition of V we have
N−1
¸
i=k
f
0
(i, x(i), u(i)) + Φ(x(N)) = f
0
(k, x, u(k))

N
¸
i=k+1
f
0
(i, x(i), u(i)) + Φ(x(N)) ≤ f
0
(k, x, u, (k)) +V (k + 1, f(k, x, u(k))¦ ,
with equality if and only if u(k +1), . . . , u(N −1) is optimal for (9.2)
k+1,x(k+1)
. Combining these
two facts we get
f
0
(k, xu

(k)) +V (k + 1, f(k, x, u

(k)))
≥ f
0
(k, x, u(k)) +V (k + 1, f(x, k, u(k))) ,
for all u(k) ∈ Ω
k
, which is equivalent to (9.4).(end theorem)
Corollary 1: Let u(k), . . . , u(N − 1) be any control for the problem (9.2)
k,x
and let x(k) =
x, . . . , x(N) be the corresponding trajectory. Then
V (, x()) ≤ f
0
(, x(), u()) +V ( + 1, f(, x(), u()), k ≤ ≤ N −1 ,
and equality holds for all k ≤ ≤ N −1 if and only if the control is optimal for (9.2)
k,x
.
Corollary 2: For k = 0, 1, . . . , N −1, let ψ(k, ) : X → Ω
k
be such that
f
0
(k, x, ψ(k, x)) +V (k + 1, f(k, x, ψ(k, x))
= Max¦f
0
(k, x, u) +V (k + 1, f(k, x, u))[u ∈ Ω
k
¦ .
Then ψ(k, ), k = 0, . . . , N − 1 is an optimal feedback control, i.e., for any k, x the control
u

(k), . . . , u

(N −1) deﬁned by u

() = ψ(, x

()), k ≤ ≤ N −1, where
124 CHAPTER 9. DYNAMIC PROGRAMING
x

( + 1) = f(, x

(), ψ(, x

()), k ≤ ≤ N −1 , x

(k) = x ,
is optimal for (α)
k,x
.
Remark: Theorem 1 and Corollary 2 are the main results of DP. The recursion equation (9.4) al-
lows us to compute the value function, and in evaluating the maximum in (9.4) we also obtain the
optimum feedback control. Note that this feedback control is optimum for all initial conditions.
However, unless we can ﬁnd a “closed-form” analytic solution to (9.4), the DP formulation may
necessitate a prohibitive amount of computation since we would have to compute and store the val-
ues of V and ψ for all k and x. For instance, suppose n = 10 and the state-space X is a ﬁnite set
with 20 elements. Then we have to compute and store 10 20 values of V , which is a reasonable
amount. But now suppose X = R
n
and we approximate each dimension of x by 20 values. Then
for N = 10, we have to compute and store 10x(20)
n
values of V . For n = 3 this number is 80,000,
and for n = 5 it is 32,000,000, which is quite impractical for existing computers. This “curse of
dimensionality” seriously limits the applicability of DP to problems where we cannot solve (9.4)
analytically.
• Exercise 1: An instructor is preparing to lead his class for a long hike. He assumes that each
person can take up to W pounds in his knapsack. There are N possible items to choose from.
Each unit of item i weighs w
i
pounds. The instructor assigns a number U
i
> 0 for each
unit of item i. These numbers represent the relative utility of that item during the hike. How
many units of each item should be placed in each knapsack so as to maximize total utility?
Formulate this problem by DP.
9.2 Continuous-time DP
We consider a continuous-time version of (9.2):
Maximize

t
f
0
f
0
(t, x(t), u(t))dt + Φ(x(t
f
))
subject to
dynamics: ˙ x(t) = f(t, x(t), u(t)) , t
0
≤ t ≤ t
f
initial condition: x(0) = x
0
,
control constraint: u : [t
0
, t
f
] → Ω and u() piecewise continuous.
(9.6)
In (9.6), x ∈ R
n
, u ∈ R
p
, Ω ⊂ R
p
. Φ : R
n
→ R is assumed differentiable and f
0
, f are assumed
to satisfy the conditions stated in VIII.1.1.
As before, for t
0
≤ t ≤ t
f
and x ∈ R
n
, let V (t, x) be the maximum value of the objective
function over the interval [t, t
f
] starting in state x at time t. Then it is easy to see that V must satisfy
V (t, x) = Max¦

t+∆
t
f
0
(τ, x(τ), u(τ))dτ
+V (t + ∆, x(t + ∆))[u : [t, t + ∆] → Ω¦, ∆ ≥ 0 ,
(9.7)
and
V (t
f
, x) = Φ(x) . (9.8)
9.2. CONTINUOUS-TIME DP 125
In (9.7), x(τ) is the solution of
˙ x(τ) = f(τ, x(τ), u(τ)) , t ≤ τ ≤ t + ∆ ,
x(t) = x .
Let us suppose that V is differentiable in t and x. Then from (9.7) we get
V (t, x) = Max¦f
0
(t, x, u)∆ +V (t, x) +
∂V
∂x
f(t, x, u)∆
+
∂V
∂t
(t, x)∆ +o(∆)[u ∈ Ω¦, ∆ > 0 .
Dividing by ∆ > 0 and letting ∆ approach zero we get the Hamilton-Jacobi- Bellman partial
differentiable equation for the value function:
∂V
∂t
(t, x) + Max¦f
0
(t, x, u) +
∂V
∂x
(t, x)f(t, x, u)[u ∈ Ω¦ = 0. (9.9)
Theorem 1: Suppose there exists a differentiable function V : [t
0
, t
f
] R
n
→ R which satisﬁes
(9.9) and the boundary condition (9.8). Suppose there exists a function ψ : [t
0
, t
f
] R
n
→ Ω
with ψ piecewise continuous in t and Lipschitz in x, satisfying
f
0
(t, x, ψ(t, x)) +
∂V
∂x
f(t, x, ψ(t, x))
= Max¦f
0
(t, x, u) +
∂V
∂x
f(t, x, u)[u ∈ Ω¦ .
(9.10)
Then ψ is an optimal feedback control for the problem (9.6), and V is the value function.
Proof: Let t ∈ [t
0
, t
f
] and x ∈ R
n
. Let ˆ u : [t, t
f
] → Ω be any piecewise continuous control and
let ˆ x(τ) be the solution of
·
ˆ x (τ) = f(τ, ˆ x(τ), ˆ u(τ)) , t ≤ τ ≤ t
f
,
ˆ x(t) = x .
(9.11)
Let x

(τ) be the solution of
˙ x

(τ) = f(τ, x

(τ), ψ(τ, x

(τ))) , t ≤ τ ≤ t
f
,
x

(τ) = x .
(9.12)
Note that the hypothesis concerning ψ guarantees a solution of (9.12). Let u

(τ) = ψ(τ, x

, (τ)), t ≤
τ ≤ t
f
. To show that ψ is an optimal feedback control we must show that

t
f
t
f
0
(tτ, x

(τ), u

(τ))dτ + Φ(x

(τ))

t
f
t
f
0
(τ, x

(τ), ˆ u(τ))dτ + Φ(ˆ x(t
f
)) .
(9.13)
To this end we note that
V (t
f
, x

(t
f
)) −V (t, x

(t)) =

t
f
f
dV

(τ, x

(τ))dτ
=∈
t
f
t
¦
∂V
∂τ
(τ, x

(τ) +
∂V
∂x
˙ x

(τ)¦dτ
= −

t
f
t
F −0(τ, x

(τ), u

(τ))dτ ,
(9.14)
126 CHAPTER 9. DYNAMIC PROGRAMING
using (9.9), (9.10). On the other hand,
V (t
f
, ˆ x(t
f
)) −V (t, ˆ x, (t)) =

t
f
t
¦
∂V
∂τ
(τ, ˆ x(τ)) +
∂V
∂x
·
˜ x (τ)¦dτ
≤ −

t
f
t
f
0
(τ, ˆ x(τ), ˆ u

(τ))dτ ,
(9.15)
using (9.9). From (9.14), (9.15), (9.8) and the fact that x

(t) = ˆ x(t) = x we conclude that
V (t, x) = Φ(x

(t
f
)) +

t
f
t
f
0
(τ, x

(τ), u

(τ))
≥ Φ(ˆ x(t
f
)) +

t
f
t
f
0
(τ, ˆ x(τ), ˆ u(τ))dτ
so that (9.13) is proved. It also follows that V is the maximum value function. ♦
• Exercise 1: Obtain the value function and the optimal feedback control for the linear regula-
tory problem:
Minimize
1
2
x

(T)P(T)x(t) +
1
2

T
0
¦x

(t)P(t)x(t)
+u

(t)Q(t)u(t)¦dt
subject to
dynamics: ˙ x(t) = A(t)x(t) +B(t)u(t) , 0 ≤ t ≤ T ,
initial condition: x(0) = x
0
,
control constraint: u(t) ∈ R
p
,
where P(t) = P

(t) is positive semi-deﬁnite, and Q(t) = Q

(t) is positive deﬁnite. [Hint:
Obtain the partial differential equation satisﬁed by V (t, x) and try a solution of the form
V (t, x) = x

R(t)x where R is unknown.]
9.3 Miscellaneous Remarks
There is vast literature dealing with the theory and applications of DP. The most elegant applications
of DP are to various problems in operations research where one can obtain “closed-form” analytic
solutions to be recursion equation for the value function. See (Bellman and Dreyfus [1952]) and
(Wagner [1969]). In the case of sequential decision-making under uncertainties DP is about the
only available general method. For an excellent introduction to this area of application see (Howard
[1960]). For an important application of DP to computational considerations for optimal control
problems see (Jacobson and Mayne [1970]). Larson [1968] has developed computational tech-
niques which greatly increase the range of applicability of DP where closed-form solutions are not
available. Finally, the book of Bellman [1957] is still excellent reading. []
Bibliography
[1] J.J. Arrow and L. Hurwicz. Decentralization and Computation in Resource in resource Allo-
cation. Essays in Economics and Econometrics. University of North Carolina Press, 1960. in
Pfouts R.W. (ed.).
[2] M. Athans and P.L. Falb. Optimal Control. McGraw-Hill, 1966.
[3] A.V. Balakrishnan and L.W. Neustadt. Computing Methods in Optimization Problems. Aca-
demic Press, 1964.
[4] K. Banerjee. Generalized Lagrange Multipliers in Dynamic Programming. PhD thesis, Col-
lege of Engineering, University of California, Berkeley, 1971.
[5] R.E. Bellman. Dynamic Programming. Princeton University Press, 1957.
[6] R.E. Bellman and S.E. Dreyfus. Applied Dynamic Programming. Princeton University Press,
1962.
[7] D. Blackwell and M.A. Girshick. Theory of Games and Statistical Decisions. John Wiley,
1954.
[8] J.E. Bruns. The function of operations research specilists in large urban schools. IEEE Trans.
on Systems Science and Cybernetics, SSC-6(4), 1970.
[9] A.E. Bryson and Y.C. Ho. Applied Optimal Control. Blaisdell, 1969.
[10] J.D. Cannon, C.D. Cullum, and E. Polak. Theory of Optimal Control and Mathematical Pro-
gramming. McGraw-Hill, 1970.
[11] G. Dantzig. Linear Programming and Extensions. Princeton University Press, 1963.
[12] C.A. Desoer. Notes for a Second Course on Linear Systems. Van Nostrand Reinhold, 1970.
[13] S.W. Director and R.A. Rohrer. On the design of resistance n-port networks by digital com-
puter. IEEE Trans. on Circuit Theory, CT-16(3), 1969a.
[14] S.W. Director and R.A. Rohrer. The generalized adjoint network and network sensitivities.
IEEE Trans. on Circuit Theory, CT-16(3), 1969b.
[15] S.W. Director and R.A. Rohrer. Automated network design–the frequency-domain case. IEEE
Trans. on Circuit Theory, CT-16(3), 1969c.
127
128 BIBLIOGRAPHY
[16] R. Dorfman and H.D. Jacoby. A Model of Public Decisions Illustrated by a Water Pollution
Policy Problem. Public Expenditures and Policy Analysis. Markham Publishing Co, 1970. in
Haveman, R.H. and Margolis, J. (eds.).
[17] R. Dorfman, P.A. Samuelson, and R.M.Solow. Linear Programming and Economic Analysis.
McGraw-Hill, 1958.
[18] R.I. Dowell and R.A. Rohrer. Automated design of biasing circuits. IEEE Trans. on Circuit
Theory, CT-18(1), 1971.
[19] Dowles Foundation Monograph. The Economic Theory of Teams. John Wiley, 1971. to appear.
[20] W.H. Fleming. Functions of Several Variables. Addison-Wesley, 1965.
[21] C.R. Frank. Production Theory and Indivisible Commodities. Princeton University Press,
1969.
[22] D. Gale. A geometric duality theorem with economic applications. Review of Economic
Studies, XXXIV(1), 1967.
[23] A.M. Geoffrion. Duality in Nonlinear Programming: a Simpliﬁed Application-Oriented Treat-
ment. The Rand Corporation, 1970a. Memo RM-6134-PR.
[24] A.M. Geoffrion. Primal resource directive approaches for optimizing nonlinear decomposable
programs. Operations Research, 18, 1970b.
[25] F.J. Gould. Extensions of lagrange multipliers in nonlinear programming. SIAM J. Apl. Math,
17, 1969.
[26] J.J. Greenberg and W.P. Pierskalla. Surrogate mathematical programming. Operations Re-
search, 18, 1970.
[27] H.J.Kushner. Introduction to Stochastic Control. Holt, Rinehart, and Winston, 1971.
[28] R.A. Howard. Dynamic Programming and Markov Processes. MIT Press, 1960.
[29] R. Isaacs. Differential Games. John Wiley, 1965.
[30] D.H. Jacobson, M.M. Lele, and J.L. Speyes. New necessary condtions of optimality for prob-
lems with state-variable inequality constraints. J. Math. Analysis and Applications, 1971. to
appear.
[31] D.H. Jacobson and D.Q. Mayne. Differential Dynamic Programming. American Elsevier
Publishing Co., 1970.
[32] S. Karlin. Mathematical Methods and Theory in Games, Programming, and Economics, vol-
[33] H.J. Kelley. Method of Gradients. Optimization Techniques. Academic Press, 1962. in Leit-
mann, G.(ed.).
BIBLIOGRAPHY 129
[34] J.H. Kelley, R.E. Kopp, and H.G. Mayer. Singular Extremals. Topics in Optimization. Aca-
demic Press, 1970. in Leitman, G. (ed.).
[35] D.A. Kendrick, H.S. Rao, and C.H. Wells. Water quality regulation with multiple polluters. In
Proc. 1971 Jt. Autom. Control Conf., Washington U., St. Louis, August 11-13 1971.
[36] T.C. Koopmans. Objectives, constraints, and outcomes in optimal growth models. Economet-
rica, 35(1), 1967.
[37] H.W. Kuhn and A.W. Tucker. Nonlinear programming. In Proc. Second Berkeley Symp. on
Math. Statistics and Probability. University of California Press, Berkeley, 1951.
[38] R.E. Larson. State Increment Dynamic Programming. American Elsevier Publishing Co.,
1968.
[39] L.S. Lasdon, S.K. Mitter, and A.D. Waren. The conjugate gradient method for optimal control
problems. IEEE Trans. on Automatic Control, AC-12(1), 1967.
[40] E.B. Lee and L. Markus. Foundation of Optimal Control Theory. John Wiley, 1967.
[41] R. Luce and H. Raiffa. Games and Decisions. John Wiley, 1957.
[42] D.G. Luenberger. Quasi-convex programming. Siam J. Applied Math, 16, 1968.
[43] O.L. Mangasarian. Nonlinear Programming. McGraw-Hill, 1969.
[44] S.R. McReynolds. The successive sweep method and dynamic programming. J. Math. Analy-
sis and Applications, 19, 1967.
[45] J.S. Meditch. Stochastic Optimal Linear Estimation and Control. McGraw-Hill, 1969.
[46] M.D. Mesarovic, D. Macho, and Y. Takahara. Theory of Hierarchical, Multi-level Systems.
[47] C.E. Miller. The Simplex Method for Local Separable Programming. Recent Advance Pro-
gramming. McGraw-Hill, 1963. in Graves, R.L. and Wolfe, P. (eds.).
[48] L.W. Neustadt. The existence of optimal controls in the absence of convexity conditions. J.
Math. Analysis and Applications, 7, 1963.
[49] L.W. Neustadt. A general theory of extremals. J. Computer and System Sciences, 3(1), 1969.
[50] H. Nikaido. Convex Structures and Economic Theory. Academic Press, 1968.
[51] G. Owen. Game Theory. W.B. Saunders & Co., 1968.
[52] E. Polak. Computational Methods in Optimization: A Uniﬁed Approach. Academic Press,
1971.
[53] L.S. Pontryagin, R.V. Boltyanski, R.V. Gamkrelidze, and E.F.Mischenko. The Mathematical
Theory of Optimal Processes. Interscience, 1962.
130 BIBLIOGRAPHY
[54] R.T. Rockafeller. Convex Analysis. Princeton University Press, 1970.
[55] M. Sakarovitch. Notes on Linear Programming. Van Nostrand Reinhold, 1971.
[56] L.J. Savage. The Foundation of Statistics. John Wiley, 1954.
[57] K. Shell. Essays in the Theory of Optimal Economic Growth. MIT Press, 1967.
[58] R.M. Solow. The economist’s approach to pollution and its control. Science, 173(3996), 1971.
[59] D.M. Topkis and A. Veinnott Jr. On the convergence of some feasible directions algorithms
for nonlinear programming. SIAM J. on Control, 5(2), 1967.
[60] H.M. Wagner. Principles of Operations Research. Prentice-Hall, 1969.
[61] P. Wolfe. The simplex method for quadratic programming. Econometrica, 27, 1959.
[62] W.M. Wonham. On the seperation theorem of optimal control. SIAM J. on Control, 6(2), 1968.
[63] E. Zangwill. Nonlinear Programming: A Uniﬁed Approach. Prentice-Hall, 1969.
Index
Active constraint, 50
augmented, 98
continuous-time, 85
augmented, 105
continuous-time, 91
discrete-time, 80
Afﬁne function, 54
Basic feasible solution, 39
basic variable, 39
Certainty-equivalence principle, 5
Complementary slackness, 34
Constraint qualiﬁcation
deﬁnition, 53
sufﬁcent conditions, 55
Continuous-time optimal control
necessary condition, 101, 103
problem formulation, 101, 103
sufﬁcient condition, 91, 125
Control of water quality, 67
Convex function
deﬁnition, 37
properties, 37, 54, 55
Convex set, 37
Derivative, 8
Design of resistive network, 15
Discrete-time optimal control
necessary condition, 78
problem formulation, 77
sufﬁcient condition, 123
Discrete-time optimality control
sufﬁcient condition, 80
Dual problem, 33, 58
Duality theorem, 33, 63
Dynamic programming, DP
optimality conditions, 123, 125
problem formulation, 121, 124
Epigraph, 61
Equilibrium of an economy, 45, 64
Farkas’ Lemma, 32
Feasible direction, 72
algorithm, 71
Feasible solution, 33, 49
Game theory , 5
Hamilton-Jacobi-Bellman equation, 125
Hamiltonian H,
˜
H, 78, 99
Hamiltonian H
˜
H, 101
Hypograph, 61
Knapsack problem, 124
Lagrange multipliers, 37
Lagrangian function, 35
Langrangian function, 21, 54
Langrangian multipliers, 21
Linear programming, LP
duality theorem, 33, 35
problem formulation, 31
theory of the ﬁrm, 42
Linear programming,LP
optimality condition, 34
Maximum principle
continuous-time, 86, 91, 101, 103
discrete-time, 80
Minimum fuel problem, 81
Minimum-time problem, 107
131
132 INDEX
example, 108
Non-degeneracy condition, 39
Nonlinear programming, NP
duality theorem, 63
necessary condition, 50, 53
problem formulation, 49
suﬁcient condition, 54
Optimal decision, 1
Optimal economic growth, 2, 113, 117
Optimal feedback control, 123, 125
Optimization over open set
necessary condition, 11
sufﬁcient condition, 13
Optimization under uncertainty, 4
Optimization with equality constraints
necessary condition, 17
sufﬁcient condition, 21
Optimum tax, 70
Primal problem, 33
optimality condition, 70
problem formulation, 70
Wolfe algorithm, 71
Recursion equation for dynamic programming,
124
Regulator problem, 81, 112
Resource allocation problem, 65
Separation theorem for convex sets, 73
Separation theorem for stochastic control, 5
Simplex algorithm, 37
Phase I, 41
Phase II, 39
Singular case for control, 113
Slack variable, 32
State-space constraint
continuous-time problem, 117
discrete-time problem, 77
Supporting hyperplane, 61, 84
Tangent, 50
Transversality condition
continuous-time problem, 91
discrete-time problem, 80
Value function, 123
Variable ﬁnal time, 103
Vertex, 38
Weak duality theorem, 33, 58
Wolfe algorithm, 71

ii

Contents
1 INTRODUCTION 2 OPTIMIZATION OVER AN OPEN SET 3 Optimization with equality constraints 4 Linear Programming 5 Nonlinear Programming 6 Discrete-time optimal control 7 Continuous-time linear optimal control 8 Coninuous-time optimal control 9 Dynamic programing 1 7 15 27 49 75 83 95 121

iii

iv CONTENTS .

Turin. The idea of making it freely available over the Web was attractive because it reafﬁrmed the original aim. However. I thank Kate Klohe for doing just that. They have urged me to re-publish it. California September. the World Wide Web has again made it possible to publish cheaply. I would appreciate knowing if you ﬁnd any mistakes in the book. 1998 P. Varaiya v . The only obstacle was to retype the manuscript in LaTex. Berkeley. However.PREFACE to this edition Notes on Optimization was published in 1971 as part of the Van Nostrand Reinhold Notes on System Sciences. or if you have suggestions for (small) changes that would improve it. Our aim was to publish short. accessible treatments of graduate-level material in inexpensive books (the price of a book in the series was about ﬁve dollars). edited by George L. Van Nostrand Reinhold was then purchased by a conglomerate which cancelled Notes on System Sciences because it was not sufﬁciently proﬁtable. Books have since become expensive.P. Notes on Optimization has been out of print for 20 years. The effort was successful for several years. several people have been using it as a text or as a reference in a course.

vi CONTENTS .

1971 P.PREFACE These Notes were developed for a ten-week course I have taught for the past three years to ﬁrst-year graduate students of the University of California at Berkeley. The treatment of the topics presented here is deep. adjoint solution) is sufﬁcient for the reader to follow the Notes. and Chapter VII before Chapter VIII. A. as well as their presentation. A reasonable knowledge of advanced calculus (up to the Implicit Function Theorem). (deterministic) optimal control. C. and mathematical economics. Billie Vrtiak for her marvelous typing in spite of starting from a not terribly legible handwritten manuscript. The selection of topics. Jacob. J-P. E.P. California November. Although the coverage is not encyclopedic. Varaiya vii . To facilitate the use of these Notes as a textbook. Chapter V must be read before Chapter VI. The examples and exercises given in the text form an integral part of the Notes and most readers will need to attend to them before continuing further. M. and Mr. basis. Finally. matrix inverse). My objective has been to present. However. Desoer. has been inﬂuenced by many of my students and colleagues. Polak. I also want to thank Mrs. I would especially like to acknowledge the help of Professors M. an understanding of this material should enable the reader to follow much of the recent technical literature on nonlinear programming. Turin for his encouraging and patient editorship. linear algebra (linear independence. and linear differential equations (transition matrix. in a compact and uniﬁed manner. the main concepts and techniques of mathematical programming and optimal control to students having diverse technical backgrounds. Ripper.A. who have read and criticized earlier drafts.L. I want to thank Professor G. Cohen. I have incurred the cost of some repetition in order to make almost all chapters self-contained. Athans. Berkeley.

viii CONTENTS .

1 The Optimal Decision Problem These Notes show how to arrive at an optimal decision assuming that complete information is given. “quality” is 1 . how much mary-john and tobacco should it purchase. The set of all decisions can be adequately represented as a subset of a vector space with each vector representing a decision. the decisions can be ranked according to whether they incur greater or lesser cost. and 2. The phrase complete information is given means that the following requirements are met: 1. The cost corresponding to these decisions is given by a real-valued function. For legal reasons the fraction α of maryjohn in the mixture must satisfy 0 < α < 1 . certain further requirements must be satisﬁed. tobacco can be purchased at a ﬁxed price. Example 1: The Pot Company (Potco) manufacturers a smoking blend called Acapulco Gold. An optimal decision is then any decision which incurs the least cost among the set of permissible decisions. When these conditions are satisﬁed. and what price p should it set? Example 2: Tough University provides “quality” education to undergraduate and graduate students. From extensive market research Potco has determined 2 their expected volume of sales as a function of α and the selling price p. 1. In order to model a decision-making situation in mathematical terms. In an agreement signed with Tough’s undergraduates and graduates (TUGs). Some illustrations will help. whereas the cost of mary-john is a function of the amount purchased. namely. we present our model of the optimal decision-making problem.Chapter 1 INTRODUCTION In this chapter. The set of all permissible decisions is known. and 2. The cost of each decision is known. illustrate decisionmaking situations by a few examples. If Potco wants to maximize its proﬁts. The blend is made up of tobacco and mary-john leaves. and brieﬂy introduce two more general models which we cannot discuss further in these Notes. Furthermore. 1.

and the remainder I(t) to investment in capital goods. one of which is a seminar and the rest of which are lecture courses. I. The current proﬁle of the ground is given by the solid line. The labor force is growing at a constant birth rate of β > 0. The University has a faculty of 1000. Subject to the agreements with the TUGs and WORs how many u’s and g’s should the President admit to maximize his rating? Example 3: (See Figure 1. Suppose that the capital stock decays exponentially with time at a rate δ > 0. The Regents of Touch rate Tough’s President at α points per u and β points per g “processed” by the University.2 CHAPTER 1. L(t)). 1] is the fraction of output which is saved and invested. b a . Figure 1. whereas each g (graduate) must take two seminars and ﬁve lecture courses. A seminar cannot have more than 20 students and a lecture course cannot have more than 40 students. each u (undergraduate) must take eight courses. There are two factors of production. The Weary Old Radicals (WORs) have a contract with the University which stipulates that every junior faculty member (there are 750 of these) shall be required to teach six lecture courses and two seminars each year. Shell is the manager of an economy which produces one output. If it costs \$c per cubic foot to excavate or ﬁll the ground. how should he design the road to meet the speciﬁcations at minimum cost? Example 4: Mr. Mr. W (t) = C(t) + I(t) = (1 − s(t))W (t) where s(t) = I(t)/W (t) . W . (Obviously. so that the net rate of growth of capital is given by the following equation: ˙ K(t) = d K(t) dt = −δK(t) + s(t)W (t) (1. .1. ∈ [0. wine. then the rate of output of wine W (t) at time is given by the production function W (t) = F (K(t). capital and labor. and K are being measured in a common currency. Hence. If K(t) and L(t) respectively are the capital stock used and the labor employed at time t.) An engineer is asked to construct a road (broken line) connection point a to point b.1: Admissable set of example.1) = −δK(t) + s(t)F (K(t). INTRODUCTION deﬁned as follows: every year. whereas every senior faculty member (there are 250 of these) shall teach three lecture courses and three seminars each year. L(t)) As Manager. Shell allocates some of the output rate W (t) to the consumption rate C(t).001. C.) Thus. The only requirement is that the ﬁnal road should not have a slope exceeding 0.

and if we let f (k) = F (k.3) where µ = (δ + β). F (λK. 0 ≤ t ≤ T .) More concisely then. w = W/L. k = K/l. The ﬁrst term of the right-hand side in (3) is the increase in the capital-to-labor ratio due to investment whereas the second terms is the decrease due to depreciation and increase in the labor force. the set of permissible decisions is represented by the set of all points in some vector space which satisfy certain constraints. then we see that F (K. (In connection with this example and related models see the critique by Koopmans [1967]. in the ﬁrst example. whence the consumption per capita of labor becomes c(t) = (l − s(t))f (k(t)).. Thus. the welfare function becomes 0 e− αt c(t)dt. these Notes are concerned with optimizing (i. the value of the mathematical exercise is greater the more insensitive are the optimum savings policies with respect to the simplifying assumptions of the mathematical model.2) it is easy to see that K(t) satisﬁes the differential equation (1. maximizing or minimizing) a real-valued function over a vector space subject to constraints. p) satisfying the constraints 0 < α < 1 and 0 < 2 p. λL) = λF (K. In the last example. any vector (u. 0 ≤ t ≤ T . Suppose there is a planning horizon time T . Shell’s savings policy s(t). . Shell starts with capital-to-labor T ratio ko . what should Mr. T ] is identiﬁed with total consumption 0 c(t)dt. In the latter case. 1) = Lf (k).3). i. constrained by 0 ≤ s(t) ≤ 1. ˙ k(t) = s(t)f (k(t)) − µk(t) (1. a permissible decision is any real-valued function s(t).1. THE OPTIMAL DECISION PROBLEM 3 ˙ L(t) = βL(t). If “welfare” over the planning period [0. a permissible decision is any two-dimensional vector (α. Example 2 is caricature (see also a faintly related but more more elaborate formulation in Bruno [1970]).2) Suppose that the production function F exhibits constant returns to scale. L)−LF (K/L. We must always remember that a mathematical formulation is inevitably an abstraction and the gain in precision may have occurred at a great loss of realism.e. whereas Example 4 is light-years away from reality. For instance.1) and (1. What is the optimum policy corresponding to this criterion? These examples illustrate the kinds of decision-making problems which can be formulated mathematically so as to be amenable to solutions by the theory presented in these Notes. g ≥ 0.e. Using these deﬁnitions and equations (1. and at time 0 Mr. be so as to maximize welfare? What savings policy maximizes welfare subject to the additional restriction that the capital-to-labor ratio at time T should be at least kT ? If future consumption is discounted at rate α > 0 and if time horizon ∞ is ∞. L) for all λ > 0. c = C/L. (It is of mathematical but not conceptual interest to note that in this case a decision is represented by a vector in a function space which is inﬁnite-dimensional. l). (1.) In the examples above. g) with u ≥ 0. The constraints themselves are presented in terms of functional inequalities or equalities.1. If we deﬁne the relevant variable in terms of per capita of labor. In the second example. constrained by the number of faculty and the agreements with the TUGs and WORs is a permissible decision.

the investor may assign probability 0. Indeed.25 that the price is unchanged. Usually there are impurities in the chemicals and disturbances in the heating process which may be regarded as additional inputs of a . A similar model is made for all the other stocks that the investor is willing to consider. How should \$1. The two objectives are incompatible. In this way. We can then redeﬁne the problem as minimizing the ﬁrst cost function (total time for trips) subject to the constraint that the waiting time for any trip is less than some reasonable bound (say one minute).2. Thus. it may be the case that the optimum decision according to the ﬁrst criterion may be lead to very long waiting times for a few trips.” More abstractly. it is important to realize that the distinction between the function which is to be optimized and the functions which describe the constraints. For instance. Before we can begin to suggest a design. Since we cannot study these more general models in these Notes. It is customary to model this uncertainty stochastically. In the ﬁrst place. He wants to maximize his capital gains. Let us suppose that we know the transportation needs of all the people in this section.1? As another example. A person wants to invest \$1. we merely point out here some situations where such models arise naturally and give some references.5 to the event that the price of shares in Glamor company increases by \$100. Now it may happen that these two objective functions may be inconsistent in the sense that they may give rise to different orderings of the permissible decisions. since the stock which is likely to have higher gains is also likely to involve greater risk. the hypothesis of complete information can be relaxed by allowing that decision-making occurs in an uncertain environment.000 be invested so as to maximize the expected value of the capital gains subject to the constraint that the probability of losing more than \$100 is less than 0.2 Some Other Models of Decision Problems Our model of a single decision-maker with complete information can be generalized along two very important directions. 1. the second goal (minimum waiting time) has been modiﬁed and reintroduced as a constraint. One way of doing this is to assign as cost to each decision the total amount of time taken to make all the trips within this section. we need a criterion by which we can compare different decisions. and a decision problem can be formulated as follows. input rates of various chemicals. although convenient for presenting the mathematical theory. etc. we need a criterion to determine what is meant by “optimum trafﬁc ﬂow. which in this case are different patterns of trafﬁc-light durations. suppose we have to choose the durations of various trafﬁc lights in a section of a city so as to achieve optimum trafﬁc ﬂow.1 Optimization under uncertainty. may be quite artiﬁcial in practice. and at the same time minimize the risk of losing his money. probability 0. We will see that in most of the results the objective function and the functions describing the constraints are treated in the same manner. we can replace the single decision-maker by a group of two or more agents whose collective decision determines the outcome. 1. The situation is different from our previous examples in that the outcome (future stock prices) is uncertain. so that this decision is far from optimum according to the second criterion. consider the design of a controller for a chemical process where the decision variable are temperature.25 that it drops by \$100. and probability 0. In the second place.000 in the stock market. INTRODUCTION At this point. An alternative and equally plausible goal may be to minimize the maximum waiting time (that is the total time spent at stop lights) in each trip.4 CHAPTER 1. This interchangeability of goal and constraints also appears at a deeper level in much of the mathematical theory.

Here there are two decision-makers with opposing objectives. then in many cases the techniques presented in these Notes can be usefully applied to the resulting optimal decision problem. it is necessary to give great attention to the various ways in which the uncertainties can be modelled mathematically. notably economics and political science. known as the Certainty-Equivalence principle in economics has been extended and baptized the Separation Theorem in the control literature. et al. an optimal decision problem under uncertainty is equivalent to another optimal decision problem under complete information. What should Alpha do to get as close to Beta as possible? What should Beta do to stay out of Alpha’s reach? This situation is fundamentally different from those discussed so far. [1969]. these problems represent one of the most important and challenging areas of research in decision theory. Each agent does not know what the other is planning to do. Although the practical impact of this theory is not great. exists which describes and prescribes behavior in these situations. SOME OTHER MODELS OF DECISION PROBLEMS 5 random nature and modeled as stochastic processes. so that optimality cannot be deﬁned as we did earlier. Situations such as these have been studied extensively and an elaborate structure. We need a new concept of rational (optimal) decision-making. it has proved to be among the most fruitful sources of unifying analytical concepts in the social sciences. since a particular decision taken by our agent may be better or worse than another decision depending upon the (unknown) decisions taken by the other agents. The difﬁculty caused by the lack of knowledge of the actions of the other decision-making agents arises even if all the agents have the same objective. To do justice to these decision-making situations.2. The best single source for Game Theory is still Luce and Raiffa [1957].1.. Kushner [1971]). For instance.2 The case of more than one decision-maker. whereas Beta is riding a motor scooter. Although problems involving many decision-makers are present in any system of large size. whereas the mathematical content of the theory is concisely displayed in Owen [1968]. After this. just as in the case of the portfolioselection problem. et al. heavy car which does not maneuver easily. The control theorist will probably be most interested in Isaacs [1965]. Agent Alpha is chasing agent Beta. however. We can only refer the reader to the extensive literature on Statistical Decision Theory (Savage [1954]. slow but with good maneuverability.2. to be able to deal with these models. The place is a large circular ﬁeld. [1970] and Marschak and Radner [1971]. it is of great significance to know that. known as the Theory of Games. (See Mesarovic. Alpha is driving a fast. Blackwell and Girshick [1954]) and on Stochastic Optimal Control (Meditch [1969]. the number of results available is pitifully small. If the uncertainties are modelled stochastically as in the example above. yet the effectiveness of his decision depends crucially upon the other’s decision. We also need to worry about ﬁnding equivalent but simpler formulations. (This result.) In the author’s opinion. and Blaquiere.) Unfortunately. 1. given appropriate conditions. we can formulate a decision problem in such a way as to take into account these random disturbances. It is of crucial importance to invent schemes to coordinate the actions of the individual decision-makers in a consistent manner.. . we need a good background in Statistics and Probability Theory besides the material presented in these Notes. See Wonham [1968].

6 CHAPTER 1. INTRODUCTION .

. . We ﬁrst establish some notation which will be in force throughout these Notes. . . . matrix multiplication. Vectors are normally denoted by lower case letters.1. yn ) then x ≥ y means xi ≥ yi . . .1 Notation 2. . with two consistent exceptions mentioned in 2. then Aj denotes the jth column of A. i = 1. If x ∈ R 2. and then we also write A = {aij }. . + xn yn as in ordinary n we deﬁne |x| = + x x. the ith component of a vector x ∈ Rn is denoted xi . . . Some additional properties are mentioned in the last section. .2 If x = (x1 . 0 denotes both the zero vector and the real number zero. . 2. 7 . if xi ≥ 0. n.1 All vectors are column vectors. . Thus if x = (x1 . .1. . . This will generalize to a canonical problem. . n. yn ) √ then x y = x1 y1 + . . and x = (x1 . . 2. . Prime denotes transpose so that if x ∈ Rn then x is the row vector x = (x1 . xn ) and y = (y1 . but no confusion will result. . its size will be clear from the context. . . xn ) . .5 below and some other minor and convenient exceptions in the text. If confusion is likely. and Ai denotes the ith row of A. . In particular if x ∈ Rn . . i = 1. . xn ). . Note that Ai is a row vector. If A is an m × n matrix. .1. we write In to denote the n × n identity matrix. then x ≥ 0. the properties of whose solution are stated as a theorem.Chapter 2 OPTIMIZATION OVER AN OPEN SET In this chapter we study in detail the ﬁrst example of Chapter 1. Then we study our example. xn ) and y = (y1 . . this entry is sometimes also denoted by the lower case letter aij .3 and 2. .1. I denotes the identity matrix. and different vectors denoted by the same symbol are distinguished by superscripts as in xj and xk . Aj denotes the entry i of A in the ith row and jth column.1. . .3 Matrices are normally denoted by capital letters.

p = sale price per pound of mixture.4 If f : Rn → Rm is a function. =  . . x x and is called the gradient of f at x. its second derivative at x is the n×n matrix (∂ 2 f /∂x∂x)(ˆ) = ˆ x fxx (ˆ) where (fxx (ˆ))j = (∂ 2 f /∂xj ∂xi )(ˆ). . then its derivative at x is the m × n matrix ˆ  x f1x (ˆ) ∂f   .6 If f : Rn → R is twice differentiable. . p) = expected sales volume (as determined by market research) of mixture as a function of(α. i = 1. ∂fm (ˆ) x ∂x1  ∂f1 x ∂xn (ˆ)     . . .1. and similarly x ˆ x ˆ x ˆ x ˆ fy (ˆ. the partial derivative of f with respect to x at the point (ˆ. If f : (x. Deﬁne the following variables and functions: α = fraction of mary-john in proposed mixture. . Note that fi : Rn → R. y ) = ((∂f /∂y1 )(ˆ. y) → f (x. .. OPTIMIZATION OVER AN OPEN SET 2.5 above.. and if the argument x x x ˆ x x is clear from the context it may be dropped. . (∂f /∂xn )(ˆ)). y )).  .2 Example We consider in detail the ﬁrst example of Chapter 1. . y ). y ) = (∂f /∂x)(ˆ. (∂f /∂ym )(ˆ. ˆ x x This derivative is denoted by (∂f /∂x)(ˆ) or fx (ˆ) or ∂f /∂x|x=ˆ or fx |x=ˆ . . x x 2. . Sometimes we describe a function by specifying a rule to calculate f (x) for every x. . . .5 If f : Rn → R is a differentiable function. . . Thus. .1. the derivative of f at x is the row vector ((∂f /∂x1 )(ˆ). y )).1. x  ∂x1. For example. v = total amount of mixture produced. ∂x fmx (ˆ) x  ∂f1 (ˆ) . . y ) = ((∂f /∂x1 )(ˆ. we can write F : x → Ax to denote the function f : Rn → Rm whose value at a point x ∈ Rn is Ax. . The column vector (fx (ˆ)) is also denoted x f (ˆ). . ∂fm x ∂xn (ˆ) 2. . 2.1. if A is an m × n matrix. y ) = (∂f /∂y)(ˆ. its ith component is written fi . y ). f (α. In this case we write f : x → f (x). . . y) is a differentiable function from ˆ Rn × Rm into R. (∂f /∂xn )(ˆ. in terms of the notation in Section 2.8 CHAPTER 2. m. . Finally. ˆ (ˆ) = fx x =  x  . y ) is the n-dimensional x ˆ row vector fx (ˆ. . . p). . x x i x fxx (ˆ) = (∂/∂x)(fx ) (ˆ). fm . if f : Rn → Rm is x ˆ x ˆ x ˆ x ˆ a differentiable function with components f1 .

EXAMPLE Since it is not proﬁtable to produce more than can be sold we must have: v = f (α. p) = P1 (αf (α.e. p∗ ) ≥ N (α. p). p) − C(α. (2. 0 < p < ∞}.2) In turn (2. p) = R(α. we 2 have the the following decision problem: Maximize N (α. so that the net proﬁt is N (α. (α∗ . h2 )) ∈ Ω for 0 ≤ δ ≤ η (2. p). p)|0 < α < 1 . p∗ ) + δ(h1 . and P2 = purchase price per pound of tobacco. i. m = αv. where Ω = {(α. Then the total cost as a function of α. The set of admissible decisions is Ω. The revenue is R(α. 9 Evidently. and t = (l − α)v.3) . p)) + P2 (1 − α)f (α. Let P1 (m) = purchase price of m pounds of mary-john. Suppose that (α∗ . h2 ) in R2 there exists η > 0 (η of course depends on h) such that ((α∗ . subject to (α. p) ∈ Ω. m = amount (in pounds) of mary-john purchased.. p). p). p) = pf (α. First of all we note that Ω is an open subset of R2 . Hence there exits ε > 0 such that (α. p) ∈ Ω. and t = amount (in pounds) of tobacco purchased. Formally. p) ∈ Ω whenever |(α. p∗) is an optimal decision. p) − (α∗ .1) We are going to establish some properties of (a∗ .2. p) for all (α. p).2) implies that for every vector h = (h1 . p∗ ). p is C(α.2. p∗ ) ∈ Ω and N (α∗ . p∗ )| < ε (2.

where oδ δ ∂N ∗ ∗ ∂p (α .9) holds for every vector h ∈ R2 .1: Admissable set of example. p∗ ) + δ(h1 .8) Thus.7) Letting δ approach zero in (2. p∗ )h1 + ∂α ∂N ∗ ∗ ∂p (α . (2. . p∗ ) is optimal. p ) = 0. p )h2 ] ∂N ∗ ∗ ∂p (α .4) yields 0 ≥ δ[ ∂N (α∗ . and δ is open.4): N (α∗ . p )h2 ] + o(δ).7). and using (2. p )h2 ] (2. we have concluded that the inequality (2. h2 ) . p∗ )h1 + ∂α Dividing by δ > 0 gives 0 ≥ [ ∂N (α∗ . OPTIMIZATION OVER AN OPEN SET 1 2 (α∗ . Ω δh (a∗ . p∗ ) | h p Figure 2. p∗ )h1 + ∂α +o(δ). Clearly this is possible only if ∂N ∗ ∗ ∂α (α . (2.6) Substitution of (2. Combining (2. using the facts that N is differentiable. p∗ )h1 + ∂α ∂N ∗ ∗ ∂p (α . p∗ + δh2 ) = +δ[ ∂N (δ∗ .8). p∗ ) N (α∗ + δh1 .6) we get 0 ≥ [ ∂N (α∗ . (2. + o(δ) δ . ∂N ∗ ∗ ∂p (α . (α∗ . p )h2 ].3) with (2. p∗ ) ≥ N (α∗ + δh1 .9) Before evaluating the usefulness of property (2. (2. p∗ + δh2 ) for 0 ≤ δ ≤ η Now we assume that the function N is differentiable so that by Taylor’s theorem N (α∗ .1) we obtain (2.5) into (2. let us prove a direct generalization.10 α CHAPTER 2.5) → 0 as δ → 0.4) (2. p ) = 0.

3. ♦ .14) yields 0 ≥ δ ∂f (x∗ )h + o(δ) ∂x and dividing by δ > 0 gives 0≥ ∂f ∗ ∂x (x )h + o(δ) δ (2.1 Theorem .3 The Main Result and its Consequences 2. Let x∗ be an optimal solution of the following decision-making problem: Maximize f (x) subject to x ∈ Ω.15) into (2.15) → 0 as δ → 0 (2. there exists ε > 0 such that x ∈ Ω whenever |x − x∗ | < ε.2.17) and taking (2.3.12) implies that for every vector h ∈ Rn there exits η > 0 (η depending on h) such that (x∗ + δh) ∈ Ω whenever 0 ≤ δ ≤ η. (2. we see that 0≥ ∂f ∗ ∂x (x )h. Then ∂f ∗ ∂x (x ) (2. (2.12) In turn.18) must hold for every h ∈ Rn .11) Proof: Since x∗ ∈ Ω and Ω is open.16) Substitution of (2. we must have 0= and the theorem is proved. Since x∗ is optimal. THE MAIN RESULT AND ITS CONSEQUENCES 11 2. Since f is differentiable. (2. by Taylor’s theorem we have f (x∗ + δh) = f (x∗ ) + where o(δ) δ ∂f ∗ ∂x (x )δh (2. (2.14) + o(δ).10) = 0. (2. Let f : Rn → R be a differentiable function. ∂f ∗ ∂x (x ).16) into account.13) (2. Let Ω be an open subset of Rn . we must then have f (x∗ ) ≥ f (x∗ + δh) whenever 0 ≤ δ ≤ η.17) Letting δ approach zero in (2.18) Since the inequality (2.

in which case (2. or.19). . the theorem does not give us any clues concerning the existence of an optimal decision.4. . p ) = ∂C ∗ ∗ ∂α (α .11) is only a necessary condition for x∗ to be optimal. OPTIMIZATION OVER AN OPEN SET Table 2.18). . x∗ ) . .4 Remarks and Extensions 2. . .1 illustrate these cases. marginal revenue = marginal cost. in these Notes we shall not be overly concerned with numerical solution techniques (but see 2.2.4.1 may x ˜ occur.12 CHAPTER 2.11) and its special case (2. Let us evaluate the usefulness of (2. p ).3. There may exist decisions x ∈ Ω ˜ such that fx (˜) = 0 but x is not optimal. Note that in the last three ﬁgures there is no optimal decision since the limit points -1 and +1 are not in the set of permissible decisions Ω = (−1.2 Consequences. p ) = ∂C ∗ ∗ ∂p (α . in the language of economic analysis. Suppose that R and C are differentiable. More generally. p∗ ) ∂R ∗ ∗ ∂α (α . 2. The diagrams in Figure 2.2 satisﬁed? Exactly one point. every optimal decision must be a solution of these n simultaneous equations of n variables. p ).18) implies that at every optimal decision (α∗ . . However. .11) gives us n equations which must be satisﬁed at any optimal decision x∗ = (x∗ .2. Equation (2.1? Yes Yes No No No Further Consequences x∗ is the unique optimal 2. n 1 These are ∂f ∗ ∂x1 (x ) = 0. say x∗ More than one point None Exactly one point More than one point Case 1 2 3 4 5 Does there exist an optimal decision for 2. We have obtained an important economic insight. ∂f ∗ ∂x2 (x ) = 0. so that the search for an optimal decision from Ω is reduced to searching among the solutions of (2.6 below). any one of the ﬁve cases in Table 2. . ∂f ∗ ∂xn (x ) =0 (2. In each case Ω = (−1. In summary. Equation (2. 1).1 A warning.1 At how many points in Ω is 2. and it does not give us sufﬁcient conditions either. 1). In practice this may be a very difﬁcult problem since these may be nonlinear equations and it may be necessary to use a digital computer. We return to the example and recall the N = R − C. ∂R ∗ ∗ ∂p (α . The theorem may also have conceptual signiﬁcance.19) Thus.

4. and by Taylor’s theorem f (x∗ + δh) = f (x∗ ) + 1 δ2 h fxx (x∗ )h + o(δ2 ). Indeed. REMARKS AND EXTENSIONS 13 -1 Case 1 1 -1 Case 2 1 -1 Case 3 1 -1 Case 4 1 -1 Case 5 1 Figure 2. Thus.20) where o(δ2 ) → 0 as δ → 0.1.2: Illustration of 4.3 Local optimum. then +1 is the optimal decision but the derivative is positive at that point. 2..11) for local optima also. This means that fxx (x∗ ) is a negative semi-deﬁnite matrix.e. Suppose f is twice-differentiable and let x∗ ∈ Ω be optimal or even locally optimal. 2 2 (2. so that dividing δ by δ2 > 0 yields 0 ≥ 1 h fxx (x∗ )h + 2 o(δ2 ) δ2 and letting δ approach zero we conclude that h fxx (x∗ )h ≤ 0 for all h ∈ Rn .2 Existence. in the third ﬁgure above.5 Sufﬁciency for local optimal.2. fx (x∗ ) = 0 and fxx is strictly negative deﬁnite. 1].4. But if Ω is closed we cannot assert that the derivative of f vanishes at the optimum.4. Now for δ > 0 sufﬁciently small f (x∗ + δh) ≤ f (x∗ ). 2. But then from the expansion (2. 2. Suppose at x∗ ∈ Ω.4. and if f is continuous. If the set of permissible decisions Ω is a closed and bounded subset of Rn . Then fx (x∗ ) = 0. We say that x∗ ∈ Ω is a locally optimal decision if there exists ε > 0 such that f (x∗ ) ≥ f (x) whenever x ∈ Ω and |x∗ − x| ≤ ε. if Ω = [−1. It is easy to see that the theorem holds (i. if we have a twice differentiable objective function.20) we can conclude that x∗ is a local optimum. 2. then it follows by the Weierstrass Theorem that there exists an optimal decision.4 Second-order conditions. . we get an additional necessary condition. 2.4.

Step 2.e. Another choice is to let di = di−1 if f (xi + di−1 x f (xi )) > f (xi ). one choice is to take di to be an optimal decision for the following problem: Max{f (xi + d x f (xi ))|d > 0. f (˜ + ε x ˜ x x f (˜)) > f (˜) for all ε > 0 sufﬁciently small. Otherwise let xi+1 = xi + di x f (xi ) and go to Step 3. Pick x0 ∈ Ω.4. This requires a one-dimensional search. Exercise: Let f be continuous differentiable.14 CHAPTER 2. stop. (xi + d x f (xi )) ∈ Ω}. i 2. For instance. Go to Step 2. Set i = i + 1 and return to Step 2. At any point x ∈ Ω the gradient x f (˜) is a direction along which f (x) increases. For other numerical procedures the reader is referred to Zangwill [1969] or Polak [1971]. We can formalize the scheme as an algorithm.11. i. Step 1. f (xi+1 ) > f (xi ) if xi+1 = xi . Set i = 0. The step size di can be selected in many ways. Calculate x f (xi ). To start the process we let d−1 > 0 be arbitrary. Then 1. Step 3. OPTIMIZATION OVER AN OPEN SET 2. otherwise let di = 1/k di−1 where k is the smallest positive integer such that f (xi + 1/k di−1 x f (xi )) > f (xi )..6 A numerical procedure. . If x f (xi ) = 0. Let {di } be produced by either of these choices and let {xi } be the resulting sequence. if x∗ ∈ Ω is a limit point of the sequence {xi }. This observation suggests the following scheme for x x ﬁnding a point x∗ ∈ Ω which satisﬁes 2. fx (x∗ ) = 0.

x∗ + ε) → V such that f1 (x. and the properties of its optimal decisions are stated in the form of a theorem.3) In particular this implies that y ∗ = g(x∗ ).2) and the decisions studied in the last chapter is that the set of permissible decisions Ω is not an open set. (ii) an open set V containing (x∗ . and that f1 (x.Chapter 3 OPTIMIZATION OVER SETS DEFINED BY EQUALITY CONSTRAINTS We ﬁrst study a simple example and examine the properties of an optimal decision. (3. y ∗ ) is an optimal decision we cannot assert that f0 (x∗ . y)|f1 (x. In the present case let 0 < ε ≤ a − x∗ and g(x) = +b[α − (x/a)2 ]1/2 . y ∗ ). g(x)) = α whenever |x − x∗ | < ε.1 (3. 1 15 . so that this assertion follows from the Implicit Function Theorem. suppose (x∗ . y) = α}. Let us suppose y ∗ = 0. y) = x2 a2 y2 b2 + = α.1) The problem can be formalized as follows (see Figure 3. Additional properties are summarized in Section 3 and a numerical scheme is applied to determine the optimal design of resistive networks. y) for all (x.1 Example We want to ﬁnd the rectangle of maximum area inscribed in an ellipse deﬁned by f1 (x. y ∗ ) is an optimal decision.1): Maximize f0 (x.1 it is evident that there exist (i)ε > 0. y) = 4xy subject to (x. Y ∗ ) = 0. y ∗ ).2) The main difference between problem (3. The assertion is false if y ∗ = 0. y) in an open set containing (x∗ . and (iii) a differentiable function g : (x∗ − ε. Since Note that y ∗ = 0 implies f1y (x∗ . y) ∈ V iff f y = g(x). y) ∈ Ω = {(x. Hence. y) = α and (x. y ∗ ) ≥ f0 (x.2). if (x∗ . Returning to problem (3. (3. 3. This will generalize to a canonical problem. Then from ﬁgure 3. Clearly then either x∗ = 0 or y ∗ = 0.

y ∗ ) = 0 we can evaluate gx (x∗ ). y ∗ ) = α and (3. (3.1: Illustration of example. (3. we see that f1x (x∗ .1. − (α/2) .3. y ∗ )gx (x∗ ) = 0. y ∗ ) + f0y (x∗ .6): −1 f0x − f0y f1y f1x = 0 at (x∗ . y ∗ ) must satisfy the two equations f1 (x∗ . g(x∗ )) is optimum for (3. (3. Solving these yields x∗ = + 1/2 a − (α/2) . and since f1y (x∗ .6) Thus an optimal decision (x∗ .2). y ∗ ) (f1x . g(x)) ≡ α for |x − x∗ | < ε. y ∗ ). y ∗ ).4) is an open set (in R1 ) and the objective function f0 is differentiable. OPTIMIZATION WITH EQUALITY CONSTRAINTS Tangent plane to Ω at (x∗ . f1y ) y∗ g(x) V ( x∗ x | ) Ω Figure 3. y ∗ ) + f1y (x∗ . ∗ ) = 0. y ∗ )gx (x∗ ) = 0 Using the fact that f1 (x.6). y ∗ ) = (x∗ . y∗ = + 1/2 b. which we can also express as ˆ so that by Theorem 2.16 CHAPTER 3.4): ˆ Maximize f0 (x) = f0 (x.5) to obtain the condition (3. −1 gx (x∗ ) = −f1y f1x (x∗ .4) ˆ But the constraint set in (3.5) and substitute in (3. it follows that x∗ is an optimal solution for (3. g(x)) subject to |x − x∗ | < ε. (x∗ . f0x (x f0x (x∗ .

y ∗ )]λ∗ . . then by re-labeling the coordinates of x if necessary. and (λ∗ ) = ∂m ∂α (3. we can assume that the m × m matrix [(∂fi /∂xj )(x∗ )]. . (3. is nonsingular. are linearly independent. . y ∗ ) = [ f1 (x∗ .12) Suppose that at x∗ the derivatives fix (x∗ ). Finally we note that λ∗ = where m(α) = maximum area.10) (3.9) is equivalent to f0 (x∗ .3. . j ≤ m. . (x∗ .6) can be interpreted differently. be continuously differentiable functions and let x∗ be an optimal decision of problem (3. Let fi : Rn → R. + 1/2 (a.9) which means that at an optimal decision the gradient of the objective function f0 is normal to the plane tangent to the constraint set Ω. (3.14) Proof. y ∗ ) = m(α) = 2αab. . . .2 General Case 3. GENERAL CASE Evidently there are two optimal decisions.11) 3. . − (α/2) 17 and the maximum area is (3. . . (3. .9): (f0x . 1. . . . Since fix (x∗ ). . b). . i = 1. are linearly independent.1 Theorem. (3. By the Implicit Function Theorem (see Fleming [1965]) it follows that there exist (i) ε > 0. f1y ) at (x∗ . i = 1. . . y ∗ ) In terms of the gradients of f0 . αm ) . Deﬁne −1 λ∗ = f0y f1y (x∗ . (ii) an .13) Furthermore.12) as a function of α = (α1 . y ∗ ). m. . . . + λ∗ fmx (x∗ ) m 1 (3.8) Then (3. m. i = 1. .8) can be rewritten as (3. let m(α1 . . If x∗ (α) is a differentiable function of α then m(α) is a differentiable function of α. αm ) be the maximum value of (3. . f1 . 1 ≤ i.2. m. . . ∂m ∂α . Then there exists a vector λ∗ = (λ∗ . Let x∗ (α) be an optimal decision for (3.12). The condition (3.2. . f0y ) = λ∗ (f1x .7) (3. i = 0. .6) and (3. λ∗ ) such that m 1 f0x (x∗ ) = λ∗ f1x (x∗ ) + . m (m < n). .12): Maximize f0 (x) subject to fi (x) = αi .

xn ) = αi .21): (f0w (x∗ ).17) ˆ But U is an open subset of Rn−m and f0 is a differentiable function on U (since f0 and g are ˆ differentiable). so that by Theorem 2. . (3. xn ) ∈ V iff xj = gj (xm+1 . . and n m+1 j fi (g(xm+1 . (3. .20) can be written as (3. . xn )]| |xm+ − x∗ | < ε. . . OPTIMIZATION WITH EQUALITY CONSTRAINTS open set V in Rn containing x∗ . f0u (u∗ ) = 0. such that m+ fi (x1 . . . . . 1 ≤ j ≤ m.3. 1 ≤ j ≤ m. xn ) .21) . In particular this implies that x∗ = gj (x∗ .18) to obtain the condition −1 −f0w fw fu + f0u = 0 at x∗ = (w∗ . and (x1 . .16) with respect to u = (xm+1 . . . . . xn ) and f = (f1 . . . .18) (3. . this is the same as f0x (x∗ ) = (λ∗ ) fx (x∗ ) = λ∗ f1x (x∗ ) + . .16) (3.2). . . . . . . (3. xn ). u∗ ) is optimal for (3.12). u) subject to u ∈ U. . m. fu (x∗ )). .19) and (3. u). = 1. . .18 CHAPTER 3. . . Differentiating (3. .15) For convenience. x∗ ). . 1 ≤ i ≤ m. + λ∗ fmx (x∗ ). and since the m × m matrix fw (x∗ ) is nonsingular we can evaluate gu (u∗ ). xn ) = αi . . . . since x∗ = (w∗ . which we can also express using the chain rule for derivatives as ˆ f0u (u∗ ) = f0w (x∗ )gu (u∗ ) + f0u (x∗ ) = 0. fm ) . Since x = (w. . let us deﬁne w = (x1 . and (xm+1 . xn ). . . where U = [(xm+1 . (3. . . xm+1 . xn ) ∈ U (see Figure 3. Then. i = 1. . .1 . deﬁne the m-dimensional column vector λ∗ by −1 (λ∗ ) = f0w fw |x∗ . . .17): ˆ Maximize f0 (u) = f0 (g(u). . . u = (xm+1 . .20) Then (3. and substitute in (3. f0u (x∗ )) = (λ∗ ) (fw (x∗ ). u∗ ) = (g(u∗ ). . we see that fw (x∗ )gu (u∗ ) + fu (x∗ ) = 0. . u∗ ). . . . n − m]. . gu (u∗ ) = −[fw (x∗)]−1 fu (x∗ ). m 1 (3. it follows that u∗ is an optimal decision for (3.19) Next. and (iii) a differentiable function g : U → Rm . . xm ) . . .

u∗ (α)). x∗ ) m 1 Ω= {x|f i (x) = αi } i = 1.23) for α ∈ N . . . .3. ∗ V xm+1 U xn 2 Figure 3. say α. . .22) (3. fw is nonsingular at m 1 m+1 ( ∗ (α). By hypothesis. . . Since f (x) and x∗ (α) are continuously differentiable by hypothesis.2.24) Differentiating (3. α so that ∗ −1 −1 wα + fw fu u∗ = fw . . . . To prove (3. xn ) (x∗ .13). We have the equation f (w∗ (α). m(α) = f0 (x∗ (α)). . . . . . . . we vary α in a neighborhood of a ﬁxed value. .x . x∗ α)) . α . it follows that f is x w nonsingular at x∗ (α) in a neighborhood of α. . . . xn ) (x∗ . −1 −f0w fw fu + f0u = 0 at (w∗ (α). . Also.2: Illustration of theorem. which is equation (3. . . x∗ (α)) and u∗ (α) = (x∗ (α). (xm+1 . . . . GENERAL CASE x1 . m .14). x m 19 g(xm+1 . so that ∗ mα = f0w wα + f0u u∗ α (3. . x∗ ) n m+1 (3. . . u∗ (α)) = α. . We deﬁne w∗ (α) = (x∗ (α). . .22) with respect to α gives ∗ fw wα + fu u∗ = I. . say N .

. and the objective function changes according to f0 (u) = f0 (g(u). u) in Ω which is not necessarily optimal. if we substitute from (3. Let us again deﬁne w = (x1 .25).25) ♦ In (3. OPTIMIZATION WITH EQUALITY CONSTRAINTS and multiplying on the left by f0w gives ∗ −1 −1 f0w wα + f0w fw fu u∗ = f0w fw . i = 1. . m}. u f0 (˜) = 0 which. together with (3. xn ) .2. Condition (3. .2 Geometric interpretation. m}.27) is equation (3. The equality constraints of the problem in 3.12 deﬁne a n − m dimensional surface Ω = {x|fi (x) = αi .27) ˜ and (3.14) and the theorem is proved. . . . .24). We shall use (3. 3. we obtain (3. . . α Using (3. α (3. so that the set of (column vectors orthogonal to this tangent surface is {λ1 x (3. . w must change according to w = g(u) (in order to ˜ ˆ maintain f (w. .28) in the last section. . in a neighborhood of x. the gradient of the objective function x f0 (x∗ ) is normal to the tangent surface (3.12). x (3.23). .13) is therefore equivalent to saying that at an optimal decision x∗ .27) ˆ Therefore. .28) ˆ u and if u is optimal.26) f1 (x∗ ) + .3 Algebraic interpretation. . u can then vary ˜ arbitrarily in a neighborhood of u. u) = α. . The hypothesis of linear independence of {fix (x∗ )|1 ≤ i ≤ m} guarantees that the tangent plane through Ω at x∗ is described by {h|fix (x∗ )h = 0 . xm ) and u = (xm+1 . u) = α). The ˆ at u is derivative of f0 ˜ ˆ u ˜ f0u (˜) = f0w gu + f0u˜ = −λ fu (˜) + f0u (˜). x x x where −1 ˜ λ = f0w fw˜ . x ˜ x (3. As u varies. 3. this equation can be rewritten as ∗ −1 f0w wα + f0u u∗ = f0w fw . . . i = 1. Suppose that fw (˜) is nonsinx gular at some point x = (w. u). i = 1. .2. . . Then the Implicit Function ˜ ˜ ˜ Theorem enables us to solve.20 CHAPTER 3. m}.13). the direction of steepest increase of f0 at u is ˜ u u f0 (˜) ˆ = −fu (˜)λ + fOu (˜) . the m equations f (w. . .20) and (3. + λm x fm (x∗ )|λi ∈ R.

13) need not hold if the derivatives fix (x∗ ). 3.2. It follows that if the functions fi .12) into a problem of maximizing f0 over an open set.3 Remarks and Extensions 3.3.3. are not linearly independent.12). λ∗ ) is a stationary point of L. Let x∗ be optimal for (3. 1 ≤ i ≤ m. Then there exists λ∗ ∈ Rm such that (x∗ .. and suppose that fix (x∗ ).3. (3. so is g (see Fleming [1965]). 1 ≤ i ≤ m. This can be checked in the following example Minimize subject to sin(x2 + x2 ) 1 2 π 2 2 2 (x1 + x2 ) = 1. and its proof is left as i=1 an exercise. The necessary condition (3. Furthermore. all ˆ . u∗ ) gu (u∗ ) = −[fw (x∗ )]−1 fu (x∗ )... Furthermore.3.16) in a neighborhood of x∗ . it is useful to translate these the comments of Section 2.12. The following is a reformulation of 3. Exercise: Show that  . 0 ≤ i ≤ m. However.3 Second-order conditions.I] where m Lww Lwu Luw Luu  gu  . λ∗ ) = 0.1. deﬁne the Lagrangian function L : Rn+m → R by L : (x. are twice continuously ˆ differentiable.4 will apply to the function f0 remarks in terms of the original function f0 and f .13) and ˆ the condition that the (n − m) × (n − m) matrix f0uu (u∗ ) is negative semi-deﬁnite. ˆ Since we can convert the problem (3.29) 3. Lx (x∗ . ∗ is a local optimum. This is possible because the function g is uniquely speciﬁed by (3.  I (w∗ .12) and (3. Keeping the notation of Theorem 3.3. are linearly independent. λ) → f0 (x) − m λi fi (x). λ∗ ) = 0 and Lλ (x∗ . ˆ f0uu (u∗ ) = [gu . then so is f0 . i. REMARKS AND EXTENSIONS 21 3. and a necessary condition for x∗ to be optimal for (3. L(x) = f0 (x) − i=1 λ∗ fi (x).e.1 The condition of linear independence. the following exercise expresses if this matrix is negative deﬁnite then x ˆ f f0uu (u∗ ) in terms of derivatives of the functions fi . if f is twice differentiable. . i .2 An alternative condition.

Then the Kirchhoff current and voltage laws respectively yield the equations Aj = 0 and A e = v (3. u)2 of the variables such that fw (xk ) is nonsingular. p ) which are under our control. vsk are the source current and voltage in the kth branch. let us suppose that there are design parameters p = (p1 .31) is replaced by (3. and g = (g1 . .6. .33) This is just a notational convenience. . Consider a network N with n + 1 nodes and b branches. (3. Find x0 arbitrary so that fi (x0 ) = αi . and f0 (uk ) = −fu (xk )λk + f0u (xk ). Here jsk . in (3. p) = g(v−v s . As before. and the ease with which we can ﬁnd wk so that f (wk .5 Design of resistive networks.31) we shall assume that gk is a function of vr . .30) as (3. uk ). x ˜ ˜ Remarks. . If f0 (uk ) = 0. Set k = 0 and go to Step 2.32) Although (3. . (3. Step 1. . Set ˜ ˜ ˜ ˜ k+1 = (wk . In the next section we apply this algorithm to a practical ˜ ˜ ˜ problem where these two steps can be carried out without too much difﬁculty.4 A numerical procedure. Hence. . jb ) respectively. We choose one of the nodes as datum and denote by e = (e1 . vs ∈ Rb for the sources. uk ) = α. Set uk = uk + dk f0 (uk ).31) where vrk is the voltage across the resistor. . uk ) = 0. Find a partition x = (w. . Orient the network graph and let v = (v1 . . The practical applicability of the algorithm depends upon two crucial factors: the ease with which we can ﬁnd a partition x = (w. thus enabling us to calculate λk . are linearly independent for all x.32): j − jx = g(vr . . . Find wk such that fi (wk . vr ∈ Rb for the resistor voltages. .30) Next we suppose that each branch k contains a (possibly nonlinear)resistive element with the form shown in Figure 3. and gk is the characteristic of the resistor. we can rewrite (3. . 1 ≤ i ≤ m. Furthermore.3. .4. . This allows us to include coupled resistors and voltagecontrolled current sources. 1 ≤ i ≤ m. Let A be the n × b reduced incidence matrix of the network graph. 2 (3. OPTIMIZATION WITH EQUALITY CONSTRAINTS 3. u) so that fw (xk ) is nonsingular. the step sizes dk > 0 can be selected various ways. so that (3. 3. . vb ) and j = (j1 . denote the vectors of branch voltages and branch currents. set k = k + 1.3. .30) implies that the current (jk −j s k) through the kth resistor depends only on the voltage vrk = (vk −v sk ) across itself. Then the following algorithm is a straightforward adaptation of the procedure in Section 2. 1 ≤ i ≤ m.3. 1 ≤ k ≤ b. stop. . so that jk − jsk = gk (vrk ) = gk (vk − vsk ). Using the obvious vector notation js ∈ Rb .22 CHAPTER 3. gb ) . en ) the vector of node-to-datum voltages. Step 2. Calculate λk −1 ˆk ˆk by (λk ) = f0w fw(xk). p). . We assume that the derivatives fix (x). Otherwise go to Step 3. The w variable may consist of any m components of x.31): j − js = g(v − vs ) = g(vr ). and return to Step 2. ˆk Step 3. no essential simpliﬁcation is achieved.

36) we see immediately that λ is the node-to-datum response voltages of a linear network N (˜r . Furthermore. then assumption (b) allows us to identify w = e. is ) subject to Ag(A e − vs . p).36) Now (3. this network has the same graph as v ˜ x the original network (since they have the same incidence matrix). G (˜r . ˜s ) be a ˜ e ˜ ˜ i ˜ is given by (see (3. for every value of (p. moreover. is ) satisfying (3. (3. vs . Also let f (x) = f (e.e.34) where we have deﬁned is = Ajs .34): Minimize f0 (e. p) − is = 0. (3.37) (3.3. p) driven by the current sources f0e (˜). p.36) has the following extremely interesting physical interpretation. p). v ˜ v ˜ .4. p. is ). (b) g is differentiable and the n×n matrix A(∂g/∂v)(v. vs . λ is the solution (unique by ˜ ˜ ˜ v ˜ v ˜ assumption (b)) of the following linear equation: AG (˜r . we have the optimization problem (3. (c) The network N described by (3. Now the crucial part in the algorithm is to obtain λk at some point xk .33) ˜ with (3. p)A . p ∈ R . The network design problem can then be stated as ﬁnding p. p) = is . vs . vs . If we combine (3. p)) v ˜ v ˜ of the original network N .3. REMARKS AND EXTENSIONS jsk jk − jsk jk o 23 + + vsk - vrk + - o vk - Figure 3. If we compare (3.3. In terms of the notation of 3. To this end let x = (˜.27)) ﬁxed point. p. is ) there is a unique e = E(p. For this reason.3: The kth branch. p) is called the adjoint network (of N ) at (˜r . vs .33): Ag(A e − vs . p. Then the corresponding λ = λ ˜ λ = f0w (˜)fw (˜) = f0e (˜)fe (˜). p)A λ = f0e (˜). To do this we make the following assumption. N (˜r . p) = (∂g/∂vr )(˜r . Assumption: (a) f0 is differentiable.35) We shall apply the algorithm 3. is ).3. x −1 x x −1 x From the deﬁnition of f we have fe (˜) = AG(˜r .32) we obtain (3. Therefore. p). its branch admittance matrix.33) is determinate i. is ) = Ag(A e−v s . vs . vs .33). v ˜ ˜ x (3. x v ˜ ˜ where vr = A e − vs . vs . is the transpose of the incremental branch admittance matrix (evaluated at (˜r . is ).. p)A is nonsingular for all v ∈ Rb .4 to this problem. is so as to minimize some speciﬁed function f0 (e. and G(˜r . if we let x = (e. p. vs . Formally. p) − is . and u = (p.29) and (3.

3. t (i) Specialize the algorithm above for this particular case. REMARKS AND EXTENSIONS f0 (e) = t∈T 25 αt |vt −v d |2 . (ii) How do the formulas change if the network equations are written using an arbitrary cutset matrix instead of the incidence matrix? .3.

26 CHAPTER 3. OPTIMIZATION WITH EQUALITY CONSTRAINTS .

1 The Linear Programming Problem 4. In the second section we present the duality theory for linear programming and use it to obtain some sensitivity results.1 Example. In Section 3 we present the Simplex algorithm which is the main procedure used to solve linear programming problems.1. In section 4 we apply the results of Sections 2 and 3 to study the linear programming theory of competitive economy. 000 and 5g+7u 40 ≤ 5250 or 5g + 7u ≤ 210. On the supply side of our accounting. For a detailed and readily accessible treatment of the material presented in this chapter see the companion volume in this Series (Sakarovitch [1971]). Let g and u respectively be the number of graduate and undergraduate students admitted. Recall Example 2 of Chapter I. Then the number of seminars demanded per year is 2g+u . 000 . 27 . Additional miscellaneous comments are collected in the last section. and the number of 20 lecture courses demanded per year is 5g+7u . the President must satisfy 2g+u 20 ≤ 2250 or 2g + u ≤ 45.Chapter 4 OPTIMIZATION OVER SETS DEFINED BY INEQUALITY CONSTRAINTS: LINEAR PROGRAMMING In the ﬁrst section we study in detail Example 2 of Chapter I. 4. the faculty can 40 offer 2(750) + 3(250) = 2250 seminars and 6(750) + 3(250) = 5250 lecture courses. and then we deﬁne the general linear programming problem. Because of his contractual agreements.

28 CHAPTER 4. is the hyperplane π(k) = {x|c x = k}. These hyperplanes for different values of k are parallel to one another since they have the same normal c. A2 h ≤ 0 . (4. LINEAR PROGRAMMING Since negative g or u is meaningless. Furthermore. and futhermore at x∗ the direction c points away from Ω. A4 x∗ < b4 . denote the rows of A.1) It is convenient to use a more general notation. u ≥ 0. Formally then the President faces the following decision problem: Maximize αg + βu subject to 2g + u ≤ 45.1. 000 5g + 7u ≤ 210. as k increases π(k) moves in the direction c. u ≥ 0 . 0) and let A be the 4×2 matrix   2 1  5 7  . For the situation depicted in Figure 4.1. b = (45000. The ﬁrst conclusion is the foundation of the powerful Simplex algorithm which we present in Section 3. (Obviously we are assuming in this discussion that c = 0. 1 (4.3) as an exercise. Here we pursue consequences of the second conclusion. A2 x∗ = b2 .2) Let Ai .2. From this condition we can immediately draw two very important conclusions: (i) at least one of the vertices of Ω is an optimal decision. c = (α. so that x ≤ y means xi ≤ yi for all i. We pause to formulate the generalization of (4.2)1 Maximize c x subject to Ax ≤ b .2) is given by Ω = {x|Ai x ≤ bi . u) .3) Recall the notation introduced in 1. So let x = (g. (4. since K ∗ lies “below” π ∗ .1) can be rewritten as (4. so that K ∗ is given by K ∗ = {x∗ + h|A1 h ≤ 0 . For each choice x. A2 h ≤ 0} . β) . 1 ≤ i ≤ 4. 000 g ≥ 0.) Evidently an optimal decision is any point x∗ ∈ Ω which lies on a hyperplane π(k) which is farthest along the direction c. 0. there are also the constraints g ≥ 0. Then the set Ω of all vectors x which satisfy the constraints in (4. We can rephrase this by saying that x∗ ∈ Ω is an optimal decision if and only if the plane π ∗ through x∗ does not intersect the interior of Ω. Now x∗ satisﬁes Ax x∗ = b1 . and A3 x∗ < b3 . A=  −1 0  0 −1 Then (4. the President receives the payoff c x. Since c x∗ ≥ c y for all y ∈ K ∗ we conclude that c h ≤ 0 for all h such that A1 h ≤ 0. .2.1 we can see that x∗ = Q is an optimal decision and the cone K ∗ is shown in Figure 4. 210000. 1 ≤ i ≤ 4} and is the polygon OP QR in Figure 4. the surface of constant payoff k say. and (ii) x∗ yields a higher payoff than all points in the cone K ∗ consisting of all rays starting at x∗ and passing through Ω. Therefore.

.4. 1 ≤ i ≤ k. Mathematically this means that (4. Consider the problem Maximize c x subject to Ai x ≤ bi . . and let bi . its generalization to n dimensions is a deep theorem known as Farkas’ lemma (see Section 2).1): Although this statement is intuitively obvious. be n-dimensional row vectors. A1 + λ∗ A2 . i ∈ I(x). THE LINEAR PROGRAMMING PROBLEM x2 29 - π∗ - π(k) = {x|c x = k} direction of increasing payoff k - P - Q = x∗ - c ⊥ π∗ A2 ⊥ P Q A1 ⊥ QR A3 . Returning to our problem.3) is satisﬁed if and only if there exist λ∗ ≥ 0.4) - - {x|A2 x = b2 } - x1 2 As c varies. Show that x∗ is optimal if an only if / c h ≤ 0 for all h such that Ai h ≤ 0 . We can see from our analysis that the situation is as follows (see Figure 4. n} be such that Ai (x) = bi . . be real numbers. Ai x < bi . . λ∗ ≥ 0 such that 1 2 c = λ∗ .3) is satisﬁed as long as c lies between A1 and A2 . let I(x) ⊂ {1. i ∈ I(x). 1 ≤ i ≤ k.1. Suppose x∗ satisﬁes the constraints. 1 2 (4. O A4 R {x|A1 x = b1 } Figure 4. For any x satisfying the constraints. the optimal decision will change. 1 ≤ i ≤ k .1: Ω = OP QR. i ∈ I(x∗ ). Exercise 1: Let Ai . Let c ∈ Rn . it is clear that (4. 2 .

6).5) is equivalent to (4. i = 1.2: K ∗ is the cone generated by Ω at x∗ . from the rest. 1 2 j (c) if ci < λ∗ + λ∗ a2i then x∗ = 0. These statements can be made in a more elegant way as follows: x∗ ∈ Ω is optimal iff there exists λ∗ ≥ 0 . λ∗ ≥ 1 2 1 2 0. λ∗ ≥ 0 such that 1 2 (a) ci ≤ λ∗ a1i + λ∗ a2i . LINEAR PROGRAMMING P x∗ = Q K∗ A2 c A1 A3 O R π∗ A4 Figure 4. (Here Ai = (ai1 . x∗ = Q is optimal iff c lies between A1 and A2 iff c = λ∗ A1 + λ∗ A2 for some λ∗ ≥ 0. i = 1. λ∗ ≥ 2 3 2 3 0. such that i 4 (a) c = i=1 λ∗ ai .6) . 2 2 3. Exercise 2: Show that (4. 2.5) accordingly We leave this as an exercise. 1 ≤ i ≤ 4.30 CHAPTER 4. i i (4. 2 1i i (4.) x∗ ∈ Ω is optimal iff there exist λ∗ ≥ 0 . x∗ = P is optimal iff c lies between A3 and A2 iff c = λ∗ A2 + λ∗ A3 for some λ∗ ≥ 0. ai2 ). and to reformulate (4. x∗ ∈ QP is optimal iff c lies along A2 iff c = λ∗ A2 for some λ∗ ≥ 0. 2.5) For purposes of application it is useful to separate those constraints which are of the form xi ≥ 0. j = 1. below. 2. 1 2 (b) if aj1 x∗ + aj2 x∗ < bj then x∗ = 0. 2. 1. (b) if Ai x∗ < bi then λ∗ = 0 . etc.

. . Proof. + ain xn ≥ bi .7) is of the form (4. .7. + ain xn = bi . . + 1 ≤ i ≤ m . There are two important special cases: Case I: (4. p + 1 ≤ j ≤ q. . . 1≤j≤n (4. . 1≤j≤p. Step 3: Replace each variable xj which is constrained xj ≤ 0 by a variable yj = −xj constrained yj ≥ 0 and then replace aij xj by (−aij )yj for every i and cj xj by (−cj )yj .1. . . (4.9) xj ≥ 0 Although (4. .7) can be transformed into an equivalent LP of the form (4. Step 1: Replace each inequality constraint aij xj ≥ bi by (−aij )xj ≤ (−bi ). . .8) xj ≥ 0 Case II: (4. q + 1 ≤ j ≤ n .2 Problem formulation.8): n (4. . k + 1 ≤ i ≤ . ail x1 + . + cn xn subject to ail x1 + ai2 x2 + . and xj ≥ 0 .9). Maximize c1 x1 + c2 x2 + . . + ain xn ≤ bi .9): n Maximize j=1 n cj xj aij xj = bi . THE LINEAR PROGRAMMING PROBLEM 31 4. . . j=1 subject to 1≤i≤m.7) is of the form (4.7) appears to be more general than (4. xj arbitary . aij . . l ≤ i ≤ k . . where the cj . . xj ≥ 0 . 1≤j≤n.1.8) and (4. .8). A linear programming problem (or LP in brief) is any decision problem of the form 4. (−aij )xj ≤ (−bi ). Step 2: Replace each equality constraint aij xj = bi by two inequality constraints: aij xj ≤ bi .4. Proposition: Every LP of the form (4. ail x1 + . bi are ﬁxed real numbers. j=1 subject to 1≤i≤m. . .7) Maximize j=1 n cj xj aij xj ≤ bi . such is not the case. .

♦ Proposition: Every LP of the form (4. . We begin by quoting a fundamental result. 1 ≤ i ≤ m (c) if j i λ∗ aij > cj then x∗ = 0 . Farkas’ Lemma. i j In the remaining discussion.1 Main results. (i) for all x ∈ Rn . Ax ≤ 0 implies c x ≤ 0.9) and is equivalent to the original one.7) can be transformed into an equivalent LP of the from (4. λk ≥ 0 such that c = i=1 λi Ai .1). λ ∈ Rk . and A = {aij } is a ﬁxed m × n matrix. Step 1: Replace each inequality constraint aij xj ≤ bi by the equality constraint aij xj + yi = bi where yi is an additional variable constrained yi ≥ 0. b ∈n are ﬁxed vectors. Let c ∈ Rn . ♦ 4. Let Ai . Ai x ≤ 0 for 1 ≤ i ≤ k implies c x ≤ 0.8) and is equivalent to the original one. An algebraic version of this result is sometimes more convenient. . . . 1 ≤ j ≤ n i m i=1 (b) if j=1 aij x∗ < bi then λ∗ = 0 . be n-dimensional row vectors.32 CHAPTER 4. LINEAR PROGRAMMING Step 4: Replace each variable xj which is not constrained in sign by a pair of variables yj −z j = xj constrained yj ≥ 0. . Step 4: Repeat these steps from the previous proposition. For a proof the reader is referred to (Mangasarian [1969]).) Step 3. The following statements are equivalent: (i) for all x ∈ Rn . Let c ∈ Rn be a column vector. Using this result it is possible to derive the main results following the intuitive reasoning of (4. Exercise 1: With the same hypothesis and notation of Exercise 1 in 4.10) and (4.1.17). We leave this development as two exercises and follow a more elegant but less intuitive approach. . Step 2: Replace each inequality constraint aij xj ≥ bi by the equality constraint aij xj − yi = bi where yi is an additional variable constrained by yi ≥ 0. 1 ≤ j ≤ m. i i Exercise 2: Let x∗ satisfy the constraints for problem (4. use the ﬁrst version of Farkas lemma to show that there exist λ∗ ≥ 0 for i ∈ I(x∗ ) such that λ∗ Ai = c . Evidently the new LP has the form (4. (The new variables added in these steps are called slack variables. Consider the pair of LPs (4. Farkas’ Lemma (algebraic version). λ∗ ≥ 0 such that m 1 m i∈I(x∗ ) (a) cj ≤ n i=1 λ∗ aij . Let A be a k × n matrix. Use the previous exercise to show that x∗ is optimal iff there exist λ∗ ≥ 0.11) . The following statements are equivalent.9) Proof. zj ≥ 0 and then replace aij xj by aij yj + (−aij )zj for every i and cj xj by cj yj + (−cj )zj .2 Qualitative Theory of Linear Programming 4. c ∈ Rn . 1 ≤ i ≤ k. whereas x ∈ Rn and λ ∈ Rm will be variable. . Evidently the resulting LP has the form (4. . k (ii) there exists λ1 ≥ 0. (ii) there exists λ ≥ 0.2. such that A λ = c.

µ ≤ 0.14) (4. such that Ax ≤ b. we must show that there exist x ≥ 0. (4. (4. λ ≥ 0. . 1 ≤ j ≤ n . Theorem 1: (Strong duality) Suppose Ωp = φ and Ωd = φ. Maximize c1 x1 + . 1 ≤ i ≤ m . . b−Ax ≥ 0 and λ ≥ 0 implies λ (b−Ax) ≥ 0 giving the second inequality. Aw = bθ ≤ 0 . λ ≥ 0. A point x ∈ Ωp (λ ∈ Ωd ) is said to be a feasible solution or feasible decision for the primal (dual). 1 ≤ i ≤ m xj ≥ 0 . Similarly let Ωd = {λ ∈ Rm |λ A ≥ c .10) is called the primal problem and (4. Furthermore. r ≤ 0 such that   A −c Im A b −I n     1    x y λ µ r     b  = c    0 By the algebraic version of Farkas’ Lemma. .10) Maximize λ1 b1 + . . 33 (4. + λm bm subject to λ1 a1j + . A λ ≥ c and b λ−c x ≤ 0. c x∗ = (λ∗ ) b. λ ≥ 0}. Lemma 1: (Weak duality) Let x ∈ Ωp .11) Deﬁnition: Let Ωp = {x ∈ Rn |Ax ≤ b. −w ≤ 0 . this is possible only if A ξ − cθ ≤ 0 . θ≤0 implies b ξ + c w ≤ 0. r ∈ R. λ ∈ Ωd . (4. then x∗ is optimal for (4. QUALITATIVE THEORY OF LINEAR PROGRAMMING below.e. . The next result is trivial. + λm amj ≥ cj . .2.11). x ≥ 0} be the set of all points satisfying the constraints of the primal problem. + ain xn ≤ bi .4. 1 ≤ j ≤ n λi ≥ 0 . µ ∈ Rm . .10) and λ∗ which is optimum for (4. i.13) . . By introducing slack variables y ∈ Rm . (4. Proof: Because of the Corollary 1 it is enough to prove the last statement.10) and λ∗ is optimal for (4..11). Then there exists x∗ which is optimum for (4. ♦ Corollary 1: If x∗ ∈ Ω and λ∗ ∈ Ωd such that c x∗ = (λ∗ ) b. ξ ≤ 0 . Then c x ≤ λ Ax ≤ λ b. y ≥ 0. + cn xn subject to ai1 x1 + .11) is called the dual problem. this is equivalent to the existence of x ≥ 0.12) Proof: x ≥ 0 and λ A − c ≥ 0 implies (λ A−c )x ≥ 0 giving the ﬁrst inequality.

Aw ≤ 0. Suppose x∗ ∈ Ωp is optimal. Then there exists an optimum decision for the primal LP iff Ωd = φ. (w/−θ) ∈ Ωp . −b ξ = b (−ξ) ≥ (Ax) (−ξ) = x (−A ξ) ≥ 0.16) Necessity. which is equivalent to (4. Consider the following exercise. j i (4. Then from Theorem 2. (4. and (A λ∗ − c) x∗ = 0 .) Proof: First of all we note that for x∗ ∈ Ωp . The sufﬁciency part of (i) follows from Theorem 1. Proof Because of the symmetry of the primal and dual it is enough to prove only (i). Ωp = φ. −w ≤ 0.16) holds for some x∗ ∈ Ωp . and c w > 0. so there exists x ≥ 0 such that Ax ≤ b. so that c x∗ = (λ∗ ) b. Suppose (4. θ) satisﬁes (4. c (x + θw) = c x + θc w. Theorem 3: (Optimality condition) x∗ ∈ Ωp is optimal if and only if there exists λ∗ ∈ Ωd such that m j=1 aij x∗ < bi implies λ∗ = 0 .15) is known as the condition of complementary slackness.15) is equivalent to (4. (4. λ∗ ∈ Ωd . there does not exist λ ≥ 0. λ ∈ Ωd .34 CHAPTER 4. while the second yields (A λ∗ ) x∗ = c x∗ .16) yields (λ∗ ) b = (λ∗ ) Ax∗ = (A λ∗ ) x∗ . Equivalently. so that only the necessity remains. that Ωd = φ. Evidently then. Also. so that (x + θw) ∈ Ωp . Then (ξ/θ) ∈ Ωd . θ) satisﬁes (4. ξ. Sufﬁciency. But (4. By Corollary 1. so that by Lemma 1 c w/(−θ) ≤ b ξ/θ. Exercise 3: Exhibit a pair of primal and dual problems such that neither has a feasible solution. LINEAR PROGRAMMING Case (i): Suppose (w. Ωd = φ. Case (ii): Suppose (w.13) and θ < 0. µ ≤ 0 such that   λ | A −In  − − −  = c | µ By Farkas’ Lemma there exists w ∈ Rn such that Aw ≤ 0. Hence. So that b ξ + c w ≤ 0. By hypothesis. in contradiction. (x + θw) ≥ 0. Then there exists an optimum decision for the dual LP iff Ωp = φ. By hypothesis. The ﬁrst equality in (4.14) since θ < 0. −ξ ≥ 0. there exist x ∈ Ωp . so that by Theorem 1 there exists λ∗ ∈ Ωd such that c x∗ = (λ∗ ) b. A(x + θw) ≤ b. ♦ The existence part of the above result can be strengthened. i j and m i=1 ((4. and c w ≤ (A λ) w = λ (Aw) ≤ 0. Suppose. the hypothesis that Ωp = φ is essential. Theorem 2: (i) Suppose Ωp = φ.16) is just an equivalent rearrangement of these two equalities. ♦ .16): (λ∗ ) (Ax∗ − b) = 0. but then for any θ > 0. ♦ Remark: In Theorem 2(i).15) λ∗ aij < cj implies x∗ = 0 . ξ.13) and θ = 0. We will show that sup {c x|x ∈ Ωp } = +∞. Now. (ii) Suppose Ωd = φ. By Lemma 1 we always have c x∗ ≤ (λ∗ ) Ax∗ ≤ (λ∗ ) b so that we must have c x∗ = (λ∗ ) Ax∗ = (λ∗ ) b. sup {c x|x ∈ Ωp } = +∞ so that there is no optimal decision for the primal. so that −A ξ ≥ 0. x∗ is optimal. λ∗ ∈ Ωd . w ≥ 0. Ωd = φ means there does not exist λ ≥ 0 such that A λ ≥ c.

17). λ ≥ 0}. λ∗ ) satisfying (4. .15) as in the following result. λ∗ ) ≤ L(x∗ .18) (4. 4. λ∗ ≥ 0 provided we strengthen (4.9).4. where L is deﬁned in (4. QUALITATIVE THEORY OF LINEAR PROGRAMMING 35 The conditions x∗ ∈ Ωp .10).17) is said to form a saddle-point of L over the set {x|x ∈ Rn .2 Results for problem (4. indicating how to use the results already obtained. . + ain xn = bi . We state these results as exercises.20) respectively.18).) Exercise 8: Formulate a dual for (4.17) Exercise 4: Prove Theorem 4. . . . (−A)x ≤ (−b).2. . Ωd denote the set of all x. unlike (4.19) .20) the dual. λ∗ ) ≤ L(x∗ . λ) = c x − λ (Ax − b) (4. The function L is called the Lagrangian.19). where L: Rn xRm → R is deﬁned by L(x. Again (4. 1 ≤ i ≤ m . whose proof is left as an exercise. λ) for all x ≥ 0. λ∗ ) ≤ L(x∗ .) Exercise 6: Show that x∗ ∈ Ωp is optimal iff there exists λ∗ ∈ Ωd such that m x∗ j > 0 implies i=1 λ∗ aij = cj . . λ) for all x ≥ 0. 1 ≤ j ≤ n . 1≤j≤n . A pair (x∗ .19) is called the primal and (4.7). λ∗ ) ≤ L(x∗ . x ≥ 0. i Exercise 7: x∗ ≥ 0 is optimal iff there exists λ∗ ∈ Rm such that L(x. (4. Remark.20) the λi are unrestricted in sign. and allλ ≥ 0. + λm amj ≥ cj (4.9). Minimize λ1 b1 + .2. xj ≥ 0 .20) Note that in (4. x∗ ∈ Ωd in Theorem 3 can be replaced by the weaker x∗ ≥ 0. . (4.19) by the equivalent LP: maximize c x. x ≥ 0} × {λ|λ ∈ Rm . Exercise 5: Prove Theorems 1 and 2 with Ωp and Ωd interpreted as above. We begin with a pair of LPs: Maximize c1 x1 + . subject to Ax ≤ b. . This is now of the form (4. (Hint. Replace (4. (Note that. λ ∈ Rm . We let Ωp . λ satisfying the constraints of (4. It is possible to derive analogous results for LPs of the form (4. + λm bm subject to λ1 a1j + . Apply Theorems 1 and 2. and obtain the result analogous to Exercise 5. Theorem 4: (Saddle point) x∗ ≥ 0 is optimal for the primal if and only if there exists λ∗ ≥ 0 such that L(x. + cn xn subject to ail x1 + . λ is not restricted in sign.

22) . ˆ ∂M + ˆ ˆ ∂cj (b. ε ∈ R. bi + ε. bi+1 . c ∈ Rn denote c(j. ˆ ∂M − ˆ ˆ ∂cj (b.11) or for the pair (4. Furthermore.2. b. cj−1 . c) ∈ B × C deﬁne M (b. c) = max {c x|x ∈ Ωp (b)} = min {λ b|λ ∈ Ωd (c)} . and for 1 ≤ j ≤ n. bm ) . ˆ c ∂M + ˆ ˆ ∂bi (b.10) or (4. if b. ε) = (c1 . We deﬁne in the usual way the right and left hand partial derivatives of M at a point (ˆ c) ∈ B × C b. c)} .20). . For 1 ≤ i ≤ m. . bi−1 . c ) − M (ˆ c)} . c2 .19) and (4. c) − M (ˆ −ε). . . . ˆ x ∈ Ωp (ˆ λ ∈ Ωd (ˆ) are optimal. ε ∈ R. b. the partial derivatives given above exist. . We investigate how the maximum value of (4. . c − M (ˆ c(j. 1≤i≤m. LINEAR PROGRAMMING 4. ◦ ◦ Theorem 5: At each (ˆ c) ∈B × C . . . . ˆ ∂M − ˆ ˆ ∂bi (b. and for (b. cj+1 . The matrix A will remain ﬁxed. then ˆ b). ε). cn ) . c) ◦ ◦ ˆ ≤ λi ≤ ∂M − ˆ ˆ ∂bi (b. c) = lim ε→0 ε>0 1 ˆ ˆ ε {M (b. . c) . C denote the interiors of B. c) = lim ε→0 ε>0 1 ˆ ˆ ε {M (b. b.10) and (4. . . c) (4. c(j. ε)) − M (ˆ c} .21) = lim ε→0 ε>0 1 ˆ ˆ ε {M (b(i. C respectively. −ε))} .3 Sensitivity analysis. . ˆ Let B . (4. ε) = (b1 . .19) changes as the vectors b and c change. b(i. b2 . We write Ωp (b) and Ωd (c) to denote the explicit dependence on b and c respectively. c) = lim ε→0 ε>0 1 ˆ ˆ ε {M (b. b ∈ Rm denote b(i. Let B = {b ∈ Rm |Ωp (b) = φ} and C = {c ∈ Rn |Ωd (c) = φ}. Let Ωp and Ωd be the sets of feasible solutions for the pair (4.36 CHAPTER 4. ˆ as follows: ∂M + ˆ ˆ ∂bi (b. cj + ε. .

ε1 )) − f (ˆ))} ≥ x x x x (1/δ1 ){f (ˆ(i. 1≤j≤n. (ii) f is said to be concave if −f is convex. ε)) − M (b. c) and M (b. so that b.(1/ε2 ){f (ˆ(i. for ε > 0. . ˆ 1 ˆ ˆ ˆ ˆ ε {M (b(i. δ1 )) − f (ˆ))} ≥ (1/δ2 ){f (ˆ(i. ˆ ε {M (b. for ε > 0. M (ˆ c) = c x.14). ε)) x. Exercise 8: (a) Show that Ωp . so that b.e. Remark 2: We can also show without difﬁculty that M (·. M (·. ε > 0. for ε > 0. ♦ We recall some fundamental deﬁnitions from convex analysis.1 Preliminaries We now present the celebrated Simplex algorithm for ﬁnding an optimum solution to any LP of the form (4. ˆ b(i. y ∈ X. (4. 0 ≤ θ ≤ 1 implies f (θx + (1 − θ)y) ≤ θf (x) + (1 − θ)f (y). 4. 0 ≤ θ ≤ 1 implies f (θx + (1 − θ)y) ≥ θf (x) + (1 − θ)f (y).22). By strong duality ˆ b. Show that at each point x in the interior of ˆ X. and the sets B ⊂ Rm .23) as ε → 0. c) . .4. This is useful in some computational problems. ˆ c ˆ 1 1 ˆ ˆ ˆ ˆ c ˆ ˆ ˆ ε {M (b. + cn xn subject to ail x1 + . −ε))} ≤ 1 {ˆ − c(j. The reason behind the last name will be clear in Section 4. c) ≥ xj ≥ ˆ ∂M − ˆ ˆ ∂cj (b. and f : X → R be convex. gives (4.3. {M (b.24): Maximize c1 x1 + . c ) − M (b. + ain xn = bi . c(j. Remark 3: The variables of the dual problem are called Lagrange variables or dual variables or shadow-prices. ε)) ≥ (ˆ(j.23) assuming that the partial derivatives exist. c)} 1 ˆ c) − M (ˆ −ε). ˆ b. 9 below. and M (ˆ c(j. c) ≤ λ ˆ ε). then we have equality in (4. the left and right hand partial derivatives of f exist. c)} ≥ ε {ˆ(j. (4. ε). 1 ≤ i ≤ m xj ≥ 0 . ε) − c} x = xj . (4. ε > 0. On the other hand. ˆ ˆˆ b. . C ⊂ Rn deﬁned above are convex sets. Exercise 9: Let X ⊂ Rn . ·) : C → R is convex. i. Taking limits as ε → 0. Finally.22). and x. . ˆ ≤ ≥ 1ˆ ˆ ˆ ˆ ε λ {b(i. −ε)} x = xj . (i) f is said to be convex if X is convex.) x x x x Remark 1: Clearly if (∂M/∂bi )(ˆ exists. ·) are piecewise linear (more accurately. δ2 )) − f (ˆ)}.3 The Simplex Algorithm 4. for ε 1ˆ ˆ ˆ −ε)} = λi . THE SIMPLEX ALGORITHM 37 ∂M + ˆ ˆ ∂cj (b.. y ∈ X and 0 ≤ θ ≤ 1 implies (θx + (1 − θ)y) ∈ X. the existence of the right and left partial derivatives follows from Exercises 8. 1 ˆ c) − M (ˆ c(j. M (b. linear plus constant) functions on B and C respectively. ˆ c ˆ ˆ ˆ ε ε which give (4. Deﬁnition: Let X ⊂ Rn and f : X → R. and then this result b) compares with 3. Deﬁnition: X ⊂ Rn is said to be convex if x. y ∈ X. ˆ ε λ {b − b(i. > 0. Then the result follows immediately.3.22). x.24) .23) Proof: We ﬁrst show (4. Ωd . ˆ b(i. 1 ≤ j ≤ n . (b) Show that for ﬁxed c ∈ C. c) : B → R is concave and for ﬁxed b ∈ B. (Hint: First show that for ε2 > ε1 > 0 > δ1 > δ2 . ε) − b}λi . c)} b(i. ε2 )) − f (ˆ)} ≥ (1/ε1 ){f (ˆ(i. M (ˆ c) = λ ˆ and by weak duality M (ˆ ε).

the inequality above can hold on if 0. j x∗ Aj + θ j j∈I(x∗ ) j∈I(x∗ ) Az(θ) = j∈I(x∗ ) zj (θ)Aj = γj Aj =b+θ·0=b. Proof: If {Aj |j ∈ I(x∗ )} is linearly independent.. with y. m n! Corollary 1: Ωp has at most vertices.24). . implies x = y = z. let z ∗ = x∗ and we are done. . not all zero. j∈I(x∗ ) j=1 For θ ∈ R deﬁne z(θ) ∈ Rn by zj (θ) = x∗ = θγj . z in Ωp and 0 < λ < 1. Then. .38 CHAPTER 4. and then c x∗ = c z(θ). . In the following we let Aj denote the jth column of A. We begin with a precise deﬁnition of a vertex. Then x is a vertex of Ωp iff {Aj |j ∈ I(x)} is a linearly independent set. Hence z(θ) ∈ Ωp whenever |θ| ≤ θ ∗ . Aj = (a1j . LINEAR PROGRAMMING As mentioned in 4. Exercise 1: Prove Lemma 1. Deﬁnition: For x ∈ Ωp . . Deﬁnition: x ∈ Ωp is said to be a vertex of Ωp if x = λy + (1 − λ)z. j ∈ I(x∗ ) j x∗ = 0 . j ∈ I(x∗ ) . i. Since Ωp has only ﬁnitely many vertices (see Corollary 1 below). (n − j)! Lemma 2: Let x∗ be an optimal decision of (4. Since x∗ is optimal we must have c x∗ ≥ c z(θ) = c x∗ + θ j∈I(x∗ ) cj yj for −∗ θ ≤ θ ≤ θ ∗ . amj ) . so that z(θ) is also an optimal solution for |θ| ≤ θ ∗ . But from the deﬁnition of z(θ) it is easy to see that we can pick θ0 with |θ0 | = θ ∗ such that zj (θ0 ) = x∗ +θ0 γj = 0 j for at least one j = j0 in I(x∗ ). Hence suppose {Aj |j ∈ I(x∗ )} is linearly dependent so that there exist γj .1 the algorithm rests upon the observations that if an optimal exists. it follows that z(θ) ≥ 0 when j |θ| ≤ min x∗ j |γj | j ∈ I(x∗ ) = θ ∗ say . let I(x) = {j|xj > 0}. cj γj = J∈I(x∗ ) Since θ can take on positive and negative values. Since x∗ > 0 for j ∈ I(x∗ ). The practicability of this investigation depends on the ease with which we can characterize the vertices of Ωp .e. Lemma 1: Let x ∈ Ωp . I(z(θ0 )) ⊂ I(x∗ ) − {j0 } . then at least one vertex of the feasible set Ωp is an optimal solution. such that γj Aj = 0 . we only have to investigate a ﬁnite set. This is done in Lemma 1. Then there is a vertex z ∗ of Ωp which is optimal.

A . For each j ∈ I(z k ) calculate cj − λ (z k )Aj . Let z 0 be a basic feasible solution obtained from Phase I or by any other means. j ∈ I(z) are called the non-basic variables at z. . and let j1 < j2 < . stop. Calculate [D(z k )]−1 . . k k k Step 4.2. We call λ(z) the shadow-price vector at z.A constitute I(z). . (4. Clearly. Otherwise pick any ˆ ∈ I(z k ) such that cˆ − λ (z k )Aj > 0 and go to Step 3. j ∈ I(z) .. stop. Lemma 3: Let z be a non-degenerate basic feasible solution. for all . j ∈ I(z). Proof: By Exercise 6 of Section 2. and the shadow-price vector λ (z k ) = c (z k )[D(z k )]−1 .26) holds iff λ = λ(z) and then (4.c(z k ).2 The Simplex Algorithm.26) (4. Assumption of non-degeneracy. Otherwise go to Step 4.25) (4. in a ﬁnite number of steps. then we let z ∗ = z(θ0 ) and we are done. cjm ) and deﬁne λ(z) by λ (z) = c (z)[D(z)]−1 . if {Aj |j ∈ I(z(θ0 ))} is linearly independent. Let D(z) denote the m × m non-singular matrix D(z) = [Aj1 . 4. . let c(z) denote the m-dimensional column vector c(z) = (cj1 . there is no ﬁnite optimum. because by Lemma 4 below. .e. we obtain a basic feasible solution. for . . . If all these numbers are ≤ 0. ♦ At this point we abandon the geometric term “vertex” and how to established LP terminology. Deﬁnition: A basic feasible solution z is said to be non-degenerate if I(z) has m elements. j2 . We shall discuss Phase II ﬁrst. Compute the vector ˆ k k γ k = (γj1 . j j Step 3. Otherwise we repeat the procedure above with z(θ0 ). and if not obtains another basic feasible solution with a higher value.4. and xj . because z k is optimal ˆ by Lemma 3.25). j ∈ I(z) . Deﬁnition: (i) z is said to be a basic feasible solution if z ∈ Ωp . Every basic feasible solution is non-degenerate. λ Aj ≥ cj . < jm .27) ♦ But since z is non-degenerate. . The algorithm is divided into two parts: In Phase I we determine if Ωp is empty or not. . sup {c x|x ∈ Ωp } = +∞. We will comment on it later. . i. < jm . Let I(z k ) consist of j1 < j2 < . Step 2. We make the following simplifying assumption. . Then z is optimal if and only if λ (z)A ≥ cj . in a ﬁnite number of steps we will ﬁnd an optimal decision z ∗ which is also vertex. .3. . . . are called the basic variables at z. Set k = 0 and go to Step 2. either we obtain an optimum solution or we discover that no optimum exists. Iterating on this procedure. . Evidently 0 < θ < ∞. . Phase II starts with a basic feasible solution and determines if it is optimal or not. Notation: Let z be a non-degenerate basic feasible solution. Deﬁne z k+1 by . The set I(z) is then called the basis at z. and {Aj |j ∈ I(z)} is linearly independent. THE SIMPLEX ALGORITHM 39 Again. z is optimal iff there exists λ such that λ Aj = cj . . . γjm ) = [D(z k )]−1 Aj . xj . Phase II: Step 1. If γ k ≤ 0. (4. jm ]. γj > 0}. and if not. for .3.27) is the same as (4. Compute θ = min {(zj γj )|j ∈ i(z). j ∈ I(z) .

Finally. Finally if we compare (4. sup {c x|x ∈ Ωp } = ∞.31) ˜ so that it is enough to prove that Aj is independent of {Aj |j ∈ I(z). Set k = k + 1 and return to Step 2. so that c z(θ) → ∞ as θ → ∞. j=ˆ and j ∈ I(z) . Proof: Deﬁne z(θ) by  k  zj − θγj . j =ˆ θ j  . Az(θ) = Az − θ j∈I(z) (4.29). j Lemma 5: z k+1 is a basic feasible solution and c z k+1 > c z k . k = 0. z(θ) ∈ Ωp for θ ≥ 0.28) we see that z˜ = 0. Then γ(z ∗ ) is an optimal solution of the dual of (4. we see case. Exercise 2: Prove Corollaries 2 and 3. Then from (4.30) = c z + θ{cˆ − λ j ˆ (z k )Aj }i . Hence. ♦ ˆ Corollary 2: In a ﬁnite number of steps Phase II will obtain an optimal solution or will determine that sup{c x|x ∈ Ωp } = ∞. j k+1 zj (4.24). j ∈ I(z) and j = ˆ . j j j j j hence I(z k+1 ) ⊂ (I(z) − {˜ j}) {ˆ . j = ˜ But if this is not the j}. k+1 k k k Proof: Let ˜ ∈ I(z k ) be such that γ˜ > 0 and z˜ = θγ˜ . j which is positive from Step 2. j ∈ I(z) θ . ♦ ˆ But from step 2 {cˆ − λ (z k )Aj } > 0. LINEAR PROGRAMMING  k k  zj − θγj .28) By Lemma 5 below.29) k γj Aj + θA = Az by deﬁnition of ˆ j γk. zj = 0 j First of all. k k c z(θ) = c z − θc (z )γ + θcˆ j ˆ = c z + θ{cˆ − c (z k )[D(z k )]−1 Aj } j (4. Lemma 4: If γ k ≤ 0. j ∈ I(z) zj (θ) = .40 CHAPTER 4. so that in (4. I(z k+1 ) has m elements. giving a contradiction.30) that c zk+1 − c zk = θ{cˆ − γ (z k )Aj } . since γ k ≤ 0 it follows that z(θ) ≥ 0 for θ ≥ 0. Next. we must have γ˜ j from (4. Corollary 3: Suppose Phase II terminates at an optimal basic feasible solution z ∗ . z k+1 is a basic feasible solution with c z k+1 > c z k . j=ˆ j =  k zj = 0 .31) we must have equality. j} (4. We see then that D(z k+1 ) is obtained from D(z k ) by replacing the column Aj by .28) and (4. Remark 1: By the non-degeneracy assumption.

so that these inverses can be easily computed. . . Phase I: Step I. 1 ≤ i ≤ m .A . . .A . . . . . We now describe how to obtain an initial basic feasible solution.32) involving the variables x and y: m Maximize − i=1 yi (4. matrix E = [Aj1 . ji−1 .A . . . Remark 3: By eliminating any dependent equations in (4.. . For details see (Canon. .4 is (∂f0 /∂uj )(xk ) − (λk ) (∂f /∂uj )(xk ).A . . + ain xn + yi = bi . jm ] and if . ji+1 . 1 ↑ ith column Then [D(z k+1 )]−1 = P M [D(z k )]−1 . . ˆ. j .A .A .A . by multiplying some of the equality constraints in (4.A . . . Hence at any degenerate basic feasible solution z k we can always ﬁnd I(z k ) ⊃ I(z k ) ¯ k ) has m elements and {Aj |j ∈ I(z k )} is a linearly independent set.4 is striking. . The analogous quantity in 3. . j. THE SIMPLEX ALGORITHM 41 . jk+1 . . ˜. . . . . if ˆ Aj m = =1 γj Aj . . . . The reason for this is that I(z k ) is not unique.A ]. Next. . But then in Step 4 it may turn out that θ = 0 so that k+1 = z k . . ji−1 . The basic variables at z k correspond to the variables wk and non-basic variables correspond to uk .   1 M =      −γ j1 γ˜ j           1 γ˜ j −γ jm γ˜ j 1 . . et al.3. . Remark 2: The similarity between Step 2 of Phase II and Step 2 of the algorithm in 3.24) by the LP (4. . ji−1 . . we can assume that b ≥ 0. so that we have to try various ¯ z ¯ k ) until we ﬁnd one for which θ > 0.A . . We can apply ¯ such that I(z ¯ Phase II using I(z k ) instead of I(z k ).4. . . Then [D(z k+1 )]−1 = P E −1 where the matrix P permutes the columns of D(z k+1 ) such that E = D(z k+1 )P . This net increase is due to the direct increase cj minus the indirect decrease λ (z k )Aj due to the compensating changes in the basic variables necessary to maintain feasibility. For each j ∈ I(z k ) we can interpret the number cj − λ (z k )Aj to be the net increase in the objective value per unit increase in the jth component of z k .  .32) subject to ail x1 + . In this way the non-degeneracy alternatives for I(z assumption can be eliminated. jk . .A . More precisely if D(z k ) = [Aj1 . 1 ≤ j ≤ n . jm ]. ji+1 . [1970]). . . . .3. . y i ≥ 0 . . .24) by −1 if necessary.A j . Replace the LP (4. it is easy to check that E −1 = M [D(z k )]−1 where  1 1 .24) we can guarantee that the matrix A ¯ has rank n. j . ˆ the column Aj . ..3. . 1 ≤ i ≤ m . . . ji+1 . . .A . . . xj ≥ 0 . Let E be the . jk < ˆ < jk+1 then D(z k+1 ) = [Aj1 . jm .A . . . ˆ. . .

. . .e. or computers. . crude oil. . . We think of a ﬁrm as a system which transforms input into outputs. y 0 ) = (0. .24) has a feasible solution iff y ∗ = 0. n ]. y ∗ ) m since the value of the objective function in (4. etc. . each activity can be conducted at any non-negative intensity or level. y∗ x∗ y∗ 4.32) lies between − i=1 bi and 0. then it combines (transforms) the input vector (a1j xj .. since human labor can do the same job as some machines and machines can replace other kinds of machines. (ii) Each activity combines the k inputs in ﬁxed proportions into the m outputs in ﬁxed proportions.24). (i) The transformation of inputs into outputs is organized into a ﬁnite number. Let . Step 3. . amj ) and B j = (bij .4. 141. . different combinations of inputs can be used to produce the same combination of outputs.24) has no feasible solution. We formalize it by specifying which transformation possibilities are available to the ﬁrm. We now make three basic assumptions about the ﬁrm. intermediate products such as steel. . . . i.) Within the ﬁrm. .32) starting with this solution. n ] and B be the k × n matrix B = [B 1 . . b) is a basic feasible solution of (4. It is these services which are consumed in the transformation into outputs. . or factory buildings.42 Go to step 2.A . If = 0. capital goods 3 such as machines of various kinds. and by an output vector we mean any k-dimensional vector y = (y1 . . . If = 0. . LINEAR PROGRAMMING Step 2. (4. . bkj ) so that if it is conducted at a level xj ≥ 0. p. Labor is not usually considered an output since slavery is not practiced. by Exercise 3 below. A be the m × n matrix [A1 . ofﬁce equipment. or raw cotton. ﬁnally various kinds of labor services. . amj xj ) = xj Aj into the output vector (b1j xj . . This substitutability among inputs is a fundamental concept in economics. 3 . . a2j . . however. the jth activity is characterized completely by two vectors Aj = (a1j . . Precisely. chemicals. By an input vector we mean any m-dimensional vector r = (r1 . . or textiles.1 Activity analysis of the ﬁrm. . of processes or activities.” dynamic Malthusian framework where the increase in labor is a function of the output. . (See the von Neumann model in (Nikaido [1968]). Furthermore. CHAPTER 4. Note that (x0 . yk ) with y ≥ 0. Go to Step 3. The ﬁrm’s outputs themselves may be raw materials (if it is a mining company) or intermediate products (if it is a steel mill) or capital goods (if it manufactures lathes) or ﬁnished goods (if it makes shirts or bakes cookies) which go directly to the consumer.32). it may be considered an output in a “closed. . Exercise 3: Show that (4. . . Phase II must terminate in an optimum based feasible solution (x∗ . is a basic feasible solution for (4. . Inputs are usually classiﬁed into raw materials such as iron ore. this transformation can be conducted in different ways.B It is more accurate to think of the services of capital goods rather than these goods themselves as inputs. . . There are m kinds of inputs and k kinds of outputs. rm ) with r ≥ 0. bkj xj ) = xj B j . Apply phase II to (4. say n. . .4 LP Theory of a Firm in a Competitive Economy 4. . .

. + 1 ≤ i ≤ m . . . rm . . r ∗ by buying or selling these inputs at the market price q1 . . . . pk ) of the outputs. . . whereas the pi . . . . Let us suppose that these ∗ inputs are 1. q.4.35): Maximize c x − (q ∗ ) ∆ subject to Ax ≤ r ∗ + ∆ . . ∗ The coefﬁcients of B and A are the ﬁxed technical coefﬁcients of the ﬁrm. (4. 4. + ain xn ≤ ri . and perhaps some raw materials. . r ∗ are in equilibrium iff for all ﬁxed ∆ ∈ Rm .33) has an optimal solution. + ain xn ≤ ri . + xn B n . . . . r ∗ are in equilibrium if and only if q ∗ is an optimal solution of (4. . . . Theorem 1: p∗ . . q ∗ ) and a set of input supplies ∗ ∗ r ∗ = (r1 . (4. + 1 ≤ i ≤ m . . We study this next. . say. qj are prices determined by the whole economy. . the ri are the ﬁxed shortterm supplies. . p∗ . ∗ ai1 x1 + . . . q ∗ ). x≥0. if he is maximizing the ﬁrm’s proﬁts. The decision variables are the activity levels x1 . Whether the ﬁrm will actually change these inputs will depend upon whether it is proﬁtable to do so. and the short-term input supplies r +1 . then it transforms the input vector x1 A1 + .4. M (∆) ≤ M (0) where M (∆) is the maximum value of the LP (4. xn . ai1 x1 + . 2. q . and q = (q1 . r ∗+1 .3 Long-term equilibrium behavior.33) subject to y = Bx. . . .34) Proof: Let c = B p∗ . . . . . ri ≥ 0 . . qm ) of the inputs is ﬁxed. 1 ≤ i ≤ . In the long run the supplies of the ﬁrst inputs are also variable and the ﬁrm can change these ∗ supplies from r1 . x∗ .2 Short-term behavior. the ﬁrm cannot change the amount available to it of some of the inputs such as capital equipment. q ∗ . . . By deﬁnition. .34): Minimize (r ∗ ) q subject to A q ≥ B p∗ q≥0. Which of these possible transformations will actually take place depends upon their relative proﬁtability and availability of inputs. . . xj ≥ 0. . . rm . 1 ≤ j ≤ n.4. certain kinds of labor. . n 1 4. . . whereas the supply of the remaining inputs can be varied.4. . We say that the prices (p∗ . With these assumptions we know all the transformations technically possible as soon as we specify the matrices A and B. faces the following decision problem: m Maximize p y − j= +1 q j rj (4. . Then the manager of the ﬁrm. . r ∗ . We assume that the ﬁrm is operating in a competitive economy which means that the unit prices p = (p1 . LP THEORY OF A FIRM IN A COMPETITIVE ECONOMY 43 (iii) If the ﬁrm conducts all the activities simultaneously with the jth activity at level xj ≥ 0. + xn An into the output vector x1 B 1 + . x∗ . . and they are available in the amounts r1 . and in turn this depends upon the prices p. . q ∗ . which the ﬁrm ac∗ cepts as given. rm ) are in (long-term) equilibrium if the ﬁrm has no proﬁt incentive to change r ∗ under the prices (p∗ . Under realistic conditions (4. In the short-term. .35) . 1 ≤ j ≤ n. .

e. r ∗ are in long-term equilibrium iff q ∗ is an optimum solution to the dual (namely (4. Thus. if the revenue of an activity is less than its input cost. cj is the revenue per unit level operation of the jth activity so that c x is the revenue when the n activities are operated at levels x. On the other hand if the jth activity is operated at level xj = 1.e.34) becomes the dual of (4. M (0) = (r ∗ ) q ∗ . .35) and q is feasible for (4..38) This relation between p∗ .16). in particular. if an equilibrium the optimum ith input supply ri is greater than the optimum demand for the ith input. c x − (q ∗ ) ∆ ≤ q (r ∗ = ∆) − (q ∗ ) ∆ . at the optimum. is i i=1 qi aij . + p∗ bkj .35). then = 0. the revenue of an activity operated at a positive level = input cost of that activity. In fact. (q ∗ ) (Ax∗ − r ∗ ) = 0 so that c x∗ = (q ∗ ) r ∗ .39) i. if x∗ is the optimum activity levels for (4. we can say even more.. (4. x∗ j . for q = q ∗ . i. q ∗ . then the input cost of operating at xj = 1. again from (4. at the optimum activity levels. c x − (q ∗ ) ∆ ≤ (q ∗ ) (r ∗ + ∆) − (q ∗ ) ∆ = (q ∗ ) r ∗ ♦ (4. Hence.15).. Now bij is the amount of the ith output produced by operating the 1 2 k jth activity at a unit level xj = 1. i. so that the input cost of operating the n activities at levels x is (A q ∗ ) = (q ∗ ) Ax. q ∗ .36) whenever x is feasible for (4.34)) of (4. then at the optimum it is ∗ operated at zero level.44 CHAPTER 4. (4. But from (4. r ∗ are in equilibrium iff c x − (q ∗ ) ∆ ≤ M (0) = (r ∗ ) q ∗ . Hence p∗ . i. (4. LINEAR PROGRAMMING For ∆ = 0.35) so that by the strong duality theorem. By weak duality if x is feasible for (4. cj = p∗ b1j + p∗ b2j + . If the ith input is valued at m a∗ .37) Remark 1: We have shown that (p∗ . it uses an amount aij of the ith input. q ∗ .38): Maximize c x subject to Ax ≤ r ∗ x≥0. Recall that c = B p∗ . and. total revenues = total cost of input supplies. . r ∗ has a very nice economic interpretation.e. Finally. (4. Also if m cj < i=1 ∗ qi aij .e. in equilibrium.38) then the output revenue is c x∗ and the input cost is (q ∗ ) Ax∗ ..15) we see that if x= astj > 0 then m cj = i=1 ∗ qi aij .34). From (4.

. .4.4. which means that the ownership of ω is divided among the various consumers j = 1. . capitalist economy. Next. capitalist economy. and it is worth decreasing this parameter if the reduction in total cost per unit decrease is i greater than λ∗ . LP THEORY OF A FIRM IN A COMPETITIVE ECONOMY n ∗ ri > j=1 45 aij x∗ . suppose that the managers of the various ﬁrms assume that the prices λ are not going to change for a long period of time. . More precisely. and ﬁnished products. and conversely it is proﬁtable to sell some i of the ith input at price qi if λ∗ < qi . . the jth consumer owns the vector of commodiJ ties ω(j) ≥ 0. . .. λ∗+1 .33). λh ) ≥ 0. then we see (assuming i differentiability) that it is worth increasing b∗ if the unit cost of increasing this parameter is less i than λ∗ . . λ∗ . q . suppose that (λ∗ . B) characterizing a ﬁrm we can suppose that all the h commodities are possible inputs and all the h commodities are possible outputs. Let us denote by M (∆1 . i 4. We are including in ω(j) the amount of his labor services which consumer j is willing to sell. from our previous analysis we know that the manager of the ith ﬁrm will plan to buy input supplies r(i) ≥ 0. . .19). the equilibrium price of an input which is in excess supply must be zero. Unfortunately we cannot present the details here. . . Suppose the resulting optimal dual variables are λ∗ . r ∗ + ∆ . Suppose that the market m 1 prices of inputs 1. .33). . This interpretation has wide applicability. ωh ) ≥ 0.4. . for an individual ﬁrm most of the inputs and most of the outputs will be zero. Thus λ∗ can be interpreted as the ﬁrm’s internal valuation of i i the ith input or the ﬁrm’s imputed or shadow price of the ith input. and carry out the i optimization problem. We think of the economy as a feedback process involving ﬁrms and consumers. . . At this time there exists within the economy an inventory of the various commodities which we can represent by a vector ω = (ω1 . λ∗ ) is an optimum solution of the dual of (4. in other words it must be a free good. ω is that portion of the outputs produced prior to T which have not been consumed up to T .22) that it is always proﬁtable to increase the ith input by buying some additional amount at price qi if λ∗ > qi . . . . . . such . We are assuming that this is a capitalist economy. . The proﬁt-maximizing behavior of the ﬁrm presented above is one of the two fundamental building blocks in the equilibrium theory of a competitive. Remark 2: Returning to the short-term decision problem (4. . i. Then. . The design procedure is to ﬁx these parameters at some nominal value b∗ . are q1 . we can see from (4. We shall limit ourselves to a rough sketch. which we mention brieﬂy.33) when the amounts of the inputs in ﬁxed supply are r1 + ∆1 . . By adding zero rows to the matrices (A. . . Let us suppose that there are a total of h commodities in the economy including raw materials. Now suppose that at time T the prevailing prices of the h commodities are λ = (λ1 . . . and j=1 ω(j) = ω. . r(i) ∈ Rh . . We observe the economy starting at time T .4 Long-term equilibrium of a competitive. intermediate and capital goods. where some of the coefﬁcients bi are design parameters. .10) or (4. . . J. j ∗ then qi = 0. Then if (∂M/∂∆i )|∆=0 exists. the sole purpose for making this change is that we no longer need to distinguish between prices of inputs and prices of outputs. labor. Often engineering design problems can be formulated as LPs of the form (4. . ∆ ) the optimum value of ∗ (4.e. Of course.

Now we come to the second building block of equilibrium theory. (4. (4. The second point is that there is no reason to expect that S(λ) ≥ 0. Thus. λ) . 1 ≤ i ≤ I . Chapter 13). so we write d(j. (4.41) we immediately conclude that J λ S(λ) = j=1 λ ω(j) . We know that r(i) and y(i) depend on λ. r(i)) is in long term equilibrium.40). We also recall that (see (4. from (4. λ) . λ) − i=1 r(i. .40) Now the ith manager can buy r(i) from only two sources: outputs from other ﬁrms. λ) = λ y(i. the ith manager can sell his planned output y(i) either as input supplies to other ﬁrms or to the consumers. λ). so that we explicitly write r(i. For a much more general mathematical treatment see (Nikaido [1968].5 Miscellaneous Comments . 2. (4. dh (j)) ≥ 0 so as to maximize his satisfaction subject to the constraint λ d(j) = λ ω(j). and the consumers who collectively own ω. . the net supply offered for sale to consumers is S(λ). (4. y(i. I. First of all. . 4.44) The most basic question of equilibrium theory is to determine conditions under which there exists a price vector λE such that the economy is in equilibrium.38)) λ r(i. .e. λ). S(λE ) = D(λE ). and Solow [1958].43) which also satisﬁes J λ D(λ) = j=1 λ ω(j) . and he will plan to produce an optimum amount. Unfortunately we must stop at this point since we cannot proceed further without introducing some more convex analysis and the ﬁxed point theorem. If we add up the buying plans of all the consumers we obtain the total demand J D(λ) = j=1 d(j.. . . because if such an equilibrium price λE exists. Chapter V). The value of the jth consumer’s possessions is λ ω(j). . λ). where J I i S(λ) = j=1 ω(j) + i=1 y(i. Similarly.41) We note two important facts.42) that is the value of the supply offered to consumers is equal to the value of the commodities (and labor) which they own. Here also d(j) will depend on λ. say y(i). For a simple treatment the reader is referred to (Dorfman. (4. Samuelson. λ) ≥ 0 .46 CHAPTER 4. . i. where I is the total number of ﬁrms. Here i = 1. then at that price the production plans of all the ﬁrms and the buying plan of all the consumers can be realized. LINEAR PROGRAMMING that (λ. The theory assumes that he will plan to buy a set of commodities d(j) = (d1 (j).

47) subject to Ax ≤ b. .45) into an equivalent LP. x ≥ 0 .3. Exercise 3: Construct an example of the kind (4. LP is today the single most important optimization technique. It is often the case in practical decision problems that the objective is not well-deﬁned.46) below. (ck ) x. Exercise 1: Show that (4.5. Exercise 1 will also indicate how to do Exercise 2. where ci : R → R are concave. in the sense that x∗ is optimal for (4.47). Maximize y subject to Ax ≤ b.45) is equivalent to (4. . y ∗ ) = (x∗ . where the ci are piecewise linear (but not concave).46) Maximize j=1 ci (xi ) (4. To obtain a feeling for the scope of LP we refer the reader to the book by one of the originators of LP (Dantzig [1963]). an elementary modiﬁcation of the Simplex algorithm can be given to obtain a “local” optimal decision.47): n (4. given the capabilities of modern computers. This is because many decision problems can be adequately formulated as LPs.45) This is not a LP since f0 is not linear. so that we have the decision problem. There may be a number of plausible objective functions. . It turns out however. piecewise-linear functions of the kind shown in Figure 4. the interpretation of “equivalent” is purposely left ambiguous. In our LP framework this situation can be formulated as follows.5. the Simplex method (together with its variants) is an extremely powerful technique for solving LPs involving thousands of variables. there are. . 4. . . . The constraints are given as usual by Ax ≤ b. 1 ≤ i ≤ k . the following exercise shows how to transform (4. See (Miller [1963]).5. f0 (x∗ )) is optimal for (4. k objective functions (c1 ) x.1 Some mathematical tricks. (4. In the next exercise. . x ≤ 0 . and. say. that even if the ci are not concave.4. Exercise 2: Obtain an equivalent LP for (4.46). It is reasonable then to deﬁne a single objective function f0 (x) by f0 (x) = minimum {(c1 ) x. x ≤ 0 y ≤ (ci ) x . .2 Scope of linear programming. Maximize f0 (x) subject to Ax ≤ b. and such that there is no equivalent LP.45) iff (x∗ . However. The above-given assumption of the concavity of the ci is crucial. MISCELLANEOUS COMMENTS 47 4. (c2 ) x. (ck ) x}. However. x ≥ 0.

xi Figure 4. .3: A function of the form used in Exercise 2. LINEAR PROGRAMMING ci (xi ) .48 CHAPTER 4. . .

however. From the discussion in 4. .Chapter 5 OPTIMIZATION OVER SETS DEFINED BY INEQUALITY CONSTRAINTS: NONLINEAR PROGRAMMING In many decision-making situations the assumption of linearity of the constraint inequalities in LP is quite restrictive. Section 3 is devoted to the important special case of quadratic programming.1): 49 . m. The linearity of the objective function is not restrictive as shown in the ﬁrst exercise below. are differentiable functions. The last section is devoted to computational considerations.1).2). we will not do this.1 Qualitative Theory of Nonlinear Programming 5. is equivalent to (5. x ∈ Rn .1). . . Two applications are given. Exercise 1: Show that (5. .1. The next exercise shows that we could restrict ourselves to objective functions which are linear. fi : Rn → R. As in Chapter 4. with variables y ∈ R. x ∈ Rn is said to be a feasible solution if it satisﬁes the constraints of (5. . . . and Ω ⊂ Rn is the subset of all feasible solutions. (5. Section 2 deals with Duality theory for the case where appropriate convexity conditions are satisﬁed. In Section 1 we present the general nonlinear programming problem (NP) and prove the Kuhn-Tucker theorem.1 The problem and elementary results. The general NP is a decision problem of the form: Maximize f0 (x) subject to (x) ≤ 0 .1) where x ∈ Rn . 1. 5.2 it is clear that equality constraints and sign constraints on some of the components of x can all be transformed into the form (5. i = 0. m. x∗ ∈ Ω is said to be an optimal decision or optimal solution if f0 (x∗ ) ≥ f0 (x) for x ∈ Ω.1. i = 1. .

let hm and {xmk . . x) so that the tangent cone is always nonempty. . x) is a closed subset of Rn . εmk > 0}∞ be such that xmk → x and (1/εmk )(xmk − x) → hm as k → ∞. . .e. and let I(x) ⊂ {1. we are interested in obtaining conditions which any optimal decision must satisfy. in Ω and a sequence of numbers εk . Deﬁnition: Let x be a feasible solution.1 and 4.1). Two more properties are stated below. C(Ω. k = 1. then θh ∈ C(Ω. . Exercise 2: (i) Show that C(Ω. . if h ∈ C(Ω. . Lemma 1: Suppose x∗ ∈ Ω is an optimum decision for (5. (See Figures 5. .1). k = 1. x). be such that (5. Proof: Let xk ∈ Ω.3) . 2.2) Returning to problem (5. . . Then f0x (x∗ )h ≤ 0 for all h ∈ C(Ω. and y − f0 (x) ≤ 0 . x) we made no use of the particular functional description of Ω. εk > 0. . with εk > 0 for all k such that k lim x = x . .18) in Chapter 2 and Exercise 1 of 4. x) = {h|h is an admissible direction for Ω at x}.1 and 5. . (ii) Let C(Ω. (Hint for (ii): For m = 1. x)}. . 2. The basic idea is to linearize the functions fi in a neighborhood of an optimal decision x∗ . fi (x) < 0 for i ∈ I(x). A vector h ∈ Rn is said to be an admissible direction for Ω at x if there exists a sequence xk . k→∞ k→∞ lim 1 (xk εk − x) = h . . Show that there exist subsequences {xmkm .) If we take xk = x and εk = 1 for all k.) Deﬁnition: (i) Let x ∈ Ω. 1 ≤ i ≤ m.. m} be such that fi (x) = 0 for ı ∈ I(x). (ii) Show that C(Ω. Suppose k=1 that hm → h as m → ∞.2 and compare them with Figures 4.50 CHAPTER 5. εmkm }∞ such that m=1 xmkm → x and (1/εmkm )(xmkm − x) → h as m → ∞.2.) In the deﬁnition of C(Ω. x) is called the tangent cone of Ω at x. 3. we see that 0 ∈ C(Ω. . . . .2. x) = {x + h|h ∈ C(Ω. 2. . (5. (The set I(x) is called the set of active constraints at x. Let K(Ω.1. x) is a cone. The argument parallels very closely that developed in Exercise 1 of 4. x) and θ ≥ 0. x∗ ) . NONLINEAR PROGRAMMING Maximize y subject to fi (x) ≤ 0. k = 1. The following elementary result is more interesting in this light and should be compared with (2. i.1 and Exercise 1 of 4.

using (5. {x|f1 (x) = 0} Figure 5. 1 (xk εk k→∞ k→∞ lim − x∗ ) = h . we have f0 (xk ) ≤ f0 (x∗ ).6) εk + o(|xk −x∗ |) εk . QUALITATIVE THEORY OF NONLINEAR PROGRAMMING direction of increasing payoff π(k) = {x|f0 (x) = k} Q x∗ 51 {x|f3 (x) = 0} P {x|f2 (x) = 0} Ω R . by Taylor’s theorem we have f0 (xk ) = f0 (x∗ + (xk − x∗ )) = f0 (x∗ ) + f0x (x∗ )(xk − x∗ ) + o(|xk − x∗ |) .5). Taking limits as k → ∞.4) and (5.5) k→∞ Since f0 is differentiable. (5. ♦ k→∞ lim o(|xk −x∗ |) |xk −x∗ | k→∞ lim |xk −x∗ | εk . so that 0 ≥ f0x (x∗ ) (x k −x∗ ) (5.4) Note that in particular (5. we can see that 0≥ = k→∞ lim f0x (x∗ ) (xk −x∗ ) εk + f0x (x∗ )h. Since xk ∈ Ω. (5.1: Ω = P QR k ∗ lim x = x .1.4) implies lim 1 εk |xk − x∗ | = |h| . and x∗ is optimal.5.

x2 ) = −x2 . by Taylor’s theorem we have fi (xk ) = fi (x∗ ) + fix (x∗ )(xk − x∗ ) + o(|xk − x∗ |) . and if i ∈ I(x∗ ). Following the proof of Lemma 1 we can conclude that 0 ≥ fix (x∗ )h. so that fi (xk ) ≤ fi (x∗ ). x∗ ) Figure 5. . in general the inclusion sign in (5. (5. . x2 ) = (x1 − 1)3 + x2 . satisfy (5. 0). εk > 0. x∗ ) in terms of the derivatives of the functions fi . fi (x∗ ) = 0. Then C(Ω. Lemma 2: Let x∗ ∈ Ω.7) Proof: Let h ∈ Rn and xk ∈ Ω. .2. x∗ ). x∗ ) - - - - 0 - C(Ω.4). fi (xk ) ≤ 0. Exercise 3: Let x ∈ R2 .52 CHAPTER 5. x∗ ) ⊂ {h|fix (x∗ )h ≤ 0 for all i ∈ I(x∗ )} . 2. . NONLINEAR PROGRAMMING x∗ K(Ω. and f2 (x1 . Since xk ∈ Ω.7) cannot be reversed. 2}. Show that 1 2 - - - . The basic problem that remains is to characterize the set C(Ω. The main reason for this is that the set {fix (x∗ )|i ∈ I(x∗ )} is not in general linearly independent. f1 (x1 . Then I(x∗ ) = {1. Unfortunately. Then we can apply Farkas’ Lemma just as in Exercise 1 of 4.2: C(Ω. Since fi is differentiable. x∗ ) is the tangent cone of Ω at x∗ . k = 1. ♦ Lemma 2 gives us a partial characterization of C(Ω. Let (x∗ . x∗ ) = (1.

5.1. QUALITATIVE THEORY OF NONLINEAR PROGRAMMING
C(Ω, x∗ ) = {h|fix (x∗ )h ≤ 0 , i = 1, 2, }. (Note that {f1x (x∗ ), f2x (x∗ )} is not a linearly independent set; see Lemma 4 below.)

53

5.1.2 Kuhn-Tucker Theorem.
Deﬁnition: Let x∗ ∈ Ω. We say that the constraint qualiﬁcation (CQ) is satisﬁed at x∗ if C(Ω, x) = {h|fix (x∗ )h ≤ 0 for all i ∈ I(x∗ )}, and we say that CQ is satisﬁed if CQ is satisﬁed at all x ∈ Ω. (Note that by Lemma 2 C(Ω, x) is always a subset of the right-hand side.) Compare the next result with Exercise 2 of 4.2. Theorem 1: (Kuhn and Tucker [1951]) Let x∗ be an optimum solution of (5.1), and suppose that CQ is satisﬁed at x∗ . Then there exist λ∗ ≥ 0, for i ∈ I(x∗ ), such that i f0x (x∗ ) =
i∈I(x∗ )

λ∗ fix (x∗ ) i

(5.8)

Proof: By Lemma 1 and the deﬁnition of CQ it follows that f0x (x∗ )h ≤ 0 whenever fix (x∗ )h ≤ 0 for all i ∈ I(x∗ ). By the Farkas’ Lemma of 4.2.1 it follows that there exist λ∗ ≥ 0 for i ∈ I(x∗ ) i such that (5.8) holds. ♦ In the original formulation of the decision problem we often have equality constraints of the form rj (x) = 0, which get replaced by rj (x) ≤ 0, −rj (x) ≤ 0 to give the form (5.1). It is convenient in application to separate the equality constraints from the rest. Theorem 1 can then be expressed as Theorem 2.

Theorem 2: Consider the problem (5.9). Maximize f0 (x) subject to fi (x) ≤ 0 , i = 1, . . . , m, rj (x) = 0 , j = 1, . . . , k .

(5.9)

Let x∗ be an optimum decision and suppose that CQ is satisﬁed at x∗ . Then there exist λ∗ ≥ 0, i = i 1, . . . , m, and µ∗ , j = 1, . . . , k such that j
m

f0x (x∗ ) =
i=1

λ∗ fix (x∗ ) + i

k j=1

µ∗ rjx (x∗ ) , j

(5.10)

and λ∗ = 0 whenever fi (x∗ ) < 0 . i Exercise 4: Prove Theorem 2. (5.11)

54

CHAPTER 5. NONLINEAR PROGRAMMING

An alternative form of Theorem 1 will prove useful for computational purposes (see Section 4). Theorem 3: Consider (5.9), and suppose that CQ is satisﬁed at an optimal solution x∗ . Deﬁne ψ : Rn → R by ψ(h) = max {−f0x (x∗ )h, f1 (x∗ ) + f1x (x∗ )h, . . . , fm (x∗ ) + fmx (x∗ )h} , and consider the decision problem Minimize ψ(h) subject to −ψ(h) − f0x (x∗ )h ≤ 0, −ψ(h) + fi (x∗ ) + fix (x∗ )h ≤ 0 , 1 ≤ i ≤ m −1 ≤ hi ≤ 1 , i = 1, . . . , n . Then h = 0 is an optimal solution of (5.12). Exercise 5: Prove Theorem 3. (Note that by Exercise 1 of 4.5, (5.12) can be transformed into a LP.) Remark: For problem (5.9) deﬁne the Lagrangian function L:
m k

(5.12)

(x1 , . . . , xn ; λ1 , . . . , λm ; µ1 , . . . , µk ) → f0 (x) −
i=1

λi fi (x) −
j=1

µj rj (x).

Then Theorem 2 is equivalent to the following statement: if CQ is satisﬁed and x∗ is optimal, then there exist λ∗ ≥ 0 and µ∗ such that Lx (x∗ , λ∗ , µ∗ ) = 0 and L(x∗ , λ∗ , µ∗ ) ≤ L(x∗ , λ, µ) for all λ ≥ 0, µ. There is a very important special case when the necessary conditions of Theorem 1 are also sufﬁcient. But ﬁrst we need some elementary properties of convex functions which are stated as an exercise. Some additional properties which we will use later are also collected here. Recall the deﬁnition of convex and concave functions in 4.2.3. Exercise 6: Let X ⊂ Rn be convex. Let h : X → R be a differentiable function. Then (i) h is convex iff h(y) ≥ h(x) + hx (x)(y − x) for all x, y, in X, (ii) h is concave iff h(y) ≤ h(x) + hx (x)(y − x) for all x, y in X, (iii) h is concave and convex iff h is afﬁne, i.e. h(x) ≡ α + b x for some ﬁxed α ∈ R, b ∈ Rn . Suppose that h is twice differentiable. Then (iv) h is convex iff hxx (x) is positive semideﬁnite for all x in X, (v) h is concave iff hxx (x) is negative semideﬁnite for all x in X, (vi) h is convex and concave iff hxx (x) ≡ 0. Theorem 4: (Sufﬁcient condition) In (5.1) suppose that f0 is concave and fi is convex for i = 1, . . . , m. Then (i) Ω is a convex subset of Rn , and (ii) if there exist x∗ ∈ Ω, λ∗ ≥ 0, i ∈ I(x∗ ), satisfying (5.8), then x∗ is an optimal solution of i (5.1). Proof: (i) Let y, z be in Ω so that fi (y) ≤ 0, fi (z) ≤ 0 for i = 1, . . . , m. Let 0 ≤ θ ≤ 1. Since fi is convex we have

5.1. QUALITATIVE THEORY OF NONLINEAR PROGRAMMING
fi (θy + (1 − θ)z) ≤ θfi (y) + (1 − θ)fi (z) ≤ 0 , 1 ≤ i ≤ m, so that (θy + (1 − θ)z) ∈ Ω, hence Ω is convex. (ii) Let x ∈ Ω be arbitrary. Since f0 is concave, by Exercise 6 we have f0 (x) ≤ f0 (x∗ ) + f0x (x∗ )(x − x∗ ) , so that by (5.8) f0 (x) ≤ f0 (x∗ ) +
i∈I(x∗ )

55

λ∗ fix (x∗ )(x − x∗ ) . i

(5.13)

Next, fi is convex so that again by Exercise 6, fi (x) ≥ fi (x∗ ) + fix (x∗ )(x − x∗ ) ; but fi (x) ≤ 0, and fi (x∗ ) = 0 for i ∈ I(x∗ ), so that fix (x∗ )(x − x∗ ) ≤ 0 for i ∈ I(x∗ ) . (5.14) Combining (5.14) with the fact that λ∗ ≥ 0, we conclude from (5.13) that f0 (x) ≤ f0 (x∗ ), so that i x∗ is optimal. ♦ Exercise 7: Under the hypothesis of Theorem 4, show that the subset Ω∗ of Ω, consisting of all the optimal solutions of (5.1), is a convex set. Exercise 8: A function h : X → R deﬁned on a convex set X ⊂ Rn is said to be strictly convex if h(θy + (1 − θ)z) < θh(y) + (1 − θ)h(z) whenever 0 < θ < 1 and y, z are in X with y = z. h is said to be strictly concave if −h is strictly convex. Under the hypothesis of Theorem 4, show that an optimal solution to (5.1) is unique (if it exists) if either f0 is strictly concave or if the fi , 1 ≤ i ≤ m, are strictly convex. (Hint: Show that in (5.13) we have strict inequality if x = x∗ .)

5.1.3 Sufﬁcient conditions for CQ.
As stated, it is usually impractical to verify if CQ is satisﬁed for a particular problem. In this subsection we give two conditions which guarantee CQ. These conditions can often be veriﬁed in practice. Recall that a function g : Rn → R is said to be afﬁne if g(x) ≡ α + b x for some ﬁxed α ∈ R and b ∈ Rn . We adopt the formulation (5.1) so that Ω = {x ∈ Rn |fi (x) ≤ 0 , 1 ≤ i ≤ m} . Lemma 3: Suppose x∗ ∈ Ω and suppose there exists h∗ ∈ Rn such that for each i ∈ I(x∗ ), either fix (x∗ )h∗ < 0, or fix (x∗ )h∗ = 0 and fi is afﬁne. Then CQ is satisﬁed at x∗ . Proof: Let h ∈ Rn be such that fix (x∗ )h ≤ 0 for i ∈ I(x∗ ). Let δ > 0. We will ﬁrst show that (h + δh∗ ) ∈ C(Ω, x∗ ). To this end let εk > 0, k = 1, 2, . . . , be a sequence converging to 0 and set xk = x∗ + εk (h + δh∗ ). Clearly xk converges to x∗ , and (1/εk )(xk − x∗ ) converges to (h + δh∗ ). Also for i ∈ I(x∗ ), if fix (x∗ )h < 0, then fi (xk ) = fi (x∗ ) + εk fix (x∗ )(h + δh∗ ) + o(εk |h + δh∗ |) ≤ δεk fix (x∗ )h∗ + o(εk |h + δh∗ |) < 0 for sufﬁciently large k , whereas for i ∈ I(x∗ ), if fi is afﬁne, then

15) Also. so wk = g(uk ) converges to w∗ = g(u∗ ).16) and recall that {fiw (x∗ )|i ∈ Jδ } is a basis in Rp we can conclude that (ξ + δξ ∗ ) = gu (u∗ )(η + δη ∗ ) so that (1/εk )(xk − x∗ ) converges to (h + hδh∗ ). u) = 0. it follows that h ∈ C(Ω. so that {fix (x∗ . fi (xk ) = fi (x∗ ) + fix (x∗ )(xk − x∗ ) + o(|xk − x∗ |) fi (x∗ ) + εk fix (x∗ )(h + δh∗ ) + o(εk ) + o(|xk − x∗ |). and set uk = u∗ + εk (η + δη ∗ ). and hence 0 = fiw (x∗ )gu (u∗ )(η + δη ∗ ) + fiu (x∗ )(η + δη ∗ ) . where U = {u ∈ Rn−p ||u − u∗ | < ρ}. Proof: Let h ∈ Rn be such that fix (x∗ )h ≤ 0 for all i ∈ I(x∗ ). NONLINEAR PROGRAMMING fi (xk ) = fi (x∗ ) + εk fix (x∗ )(h + δh∗ ) ≤ 0 for all k . fi (xk ) = fi (g(uk ). It remains to show that xk ∈ Ω for sufﬁciently large k. for i ∈ Jδ . 2 . Since g is differentiable. εk (η + δη ∗ )). or f (ˆ) ≤ 0 and f is afﬁne. u) ∈ V iff u ∈ U. Thus. wk = g(uk ). u∗ ). for i ∈ I(x∗ ) we have fi (x∗ ) < 0.16) If we compare (5. and w = g(u) . . h∗ = (ξ ∗ . (5.56 CHAPTER 5. Finally. First of all. . u) for u ∈ U so that 0 = fiw (x∗ )gu (u∗ ) + fiu (x∗ ). fix (x∗ )(h + δh∗ ) = 0}. such that fi (w. and (w. and {fix (x∗ )|i ∈ I(x∗ ). Then CQ is satisﬁed at x∗ . i ∈ I(x∗ ). ♦ Exercise 9: Suppose x∗ ∈ Ω and suppose there exists x ∈ Rn such that for each i ∈ I(x∗ ). k = 1. so that fi (xk ) < 0 for sufﬁciently large k. x∗ ). Let εk > 0. u∗ )|i ∈ Jδ } is linearly independent. (5.15) and (5. there exist ρ > 0. 0 = fi (g(u). Next we partition h.) ˆ Lemma 4: Suppose x∗ ∈ Ω and suppose there exists h∗ ∈ Rn such that fix (x∗ )h∗ ≤ 0 for i ∈ I(x∗ ). and so by deﬁnition (h + δh∗ ) ∈ C(Ω. it follows that (1/εk )(xk − x∗ ) converges to (gu (u∗ )(η + δη ∗ ). . Thus we have also shown that xk ∈ Ω for sufﬁciently large k. Since δ > 0 can be arbitrarily small. u). whereas for i ∈ Jδ . xk converges to x∗ . for i ∈ Jδ . either ˆ ∗ ) < 0 and f is convex. η ∗ ) corresponding to the partition of x = (w. x∗ ). (Hint: Show fi (x i i x i that h∗ = x − x∗ satisﬁes the hypothesis of Lemma 3. Let δ > 0. and since C(Ω. fix (x∗ )h∗ = 0} is a linearly independent set. x∗ ). x∗ ) is a closed set by Exercise 2. We will show that (h + δh∗ ) ∈ C(Ω. uk ) = 0. uk ). Then CQ is satisﬁed at x∗ . i ∈ Jδ . be any sequence converging to 0. Clearly Jδ ⊂ J = {i|i ∈ I(x∗ ). and a differentiable function g : U → Rp . . uk − u∗ ) = (1/εk )(g(uk ) − g(u∗ ). η + δη ∗ ). h∗ as h = (ξ. But for i ∈ Jδ we have 0 = fix (x∗ )(h + δh∗ ) = fiw (x∗ )(ξ + δξ ∗ ) + fiu (x∗ )(η + δη ∗ ) . We note that uk converges to u∗ . an open set V ⊂ Rn containing x∗ = (w∗ . By the Implicit Function Theorem. fi x(x∗ )h∗ = 0}. and ﬁnally xk = (sk . η). Now (1/εk )(xk − x∗ ) = (1/εk )(wk − w∗ . Let Jδ = {i|i ∈ I(x∗ ). consist of p elements.

17) then M (ˆ = f0 (ˆ). (5. we can conclude that fi (xk ) < 0 for sufﬁciently large k. .1 should be compared with Theorems 1 and 4 of 4. In 2.2. (h + δh∗ ) ∈ C(Ω.2.2. f0 : Rn → R is a given concave function.2.1 and the results in 4. f (x) ≤ b} = sup{f0 (x)|x ∈ Ω(b)} . are given convex functions. Let λ ∈ Rm . rjx (x∗ )h∗ = 0 for 1 ≤ j ≤ k. (5. let f = (f1 . Thus. We need to consider b) x the following problem also. 1 ≤ i ≤ m b x∈X . Hence. ♦ The next lemma applies to the formulation (5. and fix (x∗ )h∗ ≤ 0 for i ∈ I(x∗ ). 1 ≤ i ≤ m. k} is linearly independent. . fi : Rn → R. and C(Ω. Exercise 10: Prove Lemma 5 5.18) . fm ) : Rn → Rm . Lemma 5: Suppose x∗ is feasible for (5. b b b For convenience. x∗ ). fix (x∗ )h∗ = 0} {rjx (x∗ )|j = 1. However. B = {b|Ω(b) = φ}. We can only present some of the basic results here.3 we give some application of duality theory and in 2. and M : B→R {+∞} by M (b) = sup{f0 (x)|x ∈ X. so that in particular if x∗ is an optimal solution of (5. .17) where x ∈ Rn . Then CQ is satisﬁed at x∗ . x∗ ). .1 Basic results. We wish to examine the behavior of the maximum value of (5. λ ≥ 0. To ﬁnish the proof we note that δ > 0 can be made arbitrarily small. . . we will give some geometric insight.2 Duality Theory Duality theory is perhaps the most beautiful part of nonlinear programming. xk ∈ Ω for sufﬁciently large k. Consider problem (5. X is a given convex subset of Rn and ˆ = (ˆ1 .17) as ˆ varies. x∗ ) is closed by Exercise 2. DUALITY THEORY 57 and since fi (x∗ ) = 0 whereas fix (x∗ )(h + δh∗ ) < 0. . .2 we refer to some of the important generalizations. in terms of suggesting important computational algorithms. It has resulted in many applications within nonlinear programming. . f (x) ≤ b}. ˆm ) is a given vector.9). So we deﬁne b Ω(b) = {x|x ∈ X. so that h ∈ C(Ω. 5. and it has provided many unifying conceptual insights into economics and management science. It may be useful to note in the following discussion that most of the results do not require differentiability of the various functions.17) which we call the primal problem: Maximize f0 (x) subject to fi (x) ≤ ˆi . and even so some of the proofs are relegated to the Appendix at the end of this Chapter since they depend on advanced material. be ﬁxed.9) and suppose there exists h∗ ∈ Rn such that the set {fix (x∗ )|i ∈ I(x∗ ). The results in 2. . . Maximize f0 (x) − λ (f (x) − ˆ b) subject to x ∈ X .3.5. Its proof is left as an exercise since it is very similar to the proof of Lemma 4. .

19) is called the dual problem: Minimize m(λ) subject to λ ≥ 0 . NONLINEAR PROGRAMMING m(λ) = sub{f0 (x) − λ (f (x) − ˆ ∈ X} . b) b).19) become simple..17) is usually equal to Rn and then. Lemma 1 shows that the cost function of the dual problem is convex which is useful information since there are computation techniques which apply to convex cost functions but not to arbitrary nonlinear cost functions. λ) with x ∈ X.2 below.3.3. b) Proof: Since f (x) − ˆ ≤ 0. we have λ (f (x) − ˆ ≤ 0. Remark 1: The set X in (5. f. x ∈ Ω(ˆ and if λ ≥ 0. (Here R+ = {λ ∈ Rn |λ ≥ 0}. there is no reason to separate it out. i. So. and λ ≥ 0. For example see the problems discussed in Sections 2. n n Lemma 1: m : R+ → R {+∞} is a convex function. we have f0 (x) ≤ M (ˆ ≤ m(λ) for x ∈ Ω(ˆ λ ≥ 0 . We ﬁrst give a simple sufﬁciency condition.18) and the solution of the dual problem (5.19) Let m∗ = inf {m(λ)|λ ≥ 0}. f0 (x) ≤ M (ˆ ≤ m∗ ≤ m(λ) .1 and 2. (5. Lemma 2: (Weak duality) If x is feasible for (5. ♦ ˆ = m∗ in The basic problem of Duality Theory is to determine conditions under which M (b) (5. ˆ Deﬁnition: A pair (ˆ. then b).20). X. of course. and since M (ˆ is independent of λ. b). and λ ≤ 0 is said to satisfy the optimality conditions if x ˆ ˆ (5. Remark 2: It is sometimes useful to know that Lemmas 1 and 2 below hold without any convexity conditions on f0 . ≤ sup {f0 (x) − λ (f (x) − b)|x b)} b) ≤ sup {f0 (x) − λ (f (x) − ˆ ∈ X} = m(λ) . b)|x Problem (5.20) .) Exercise 1: Prove Lemma 1. b b) f0 (x) ≤ f0 (x) − λ (f (x) − ˆ for x ∈ Ω(ˆ λ ≥ 0 .58 and deﬁne CHAPTER 5. if we take the inﬁmum with respect to λ ≥ 0 in the right-hand b) inequality we get (5. Hence f0 (x) ≤ sup {f0 (x)|x ∈ Ω(ˆ = M (ˆ b)} b) ˆ ∈ Ω(ˆ and since Ω(ˆ ⊂ X. b)|x Thus.17).e. b). it is sometimes possible to include some of the constraints in X in such a way that the calculation of m(λ) by (5. Lemma 2 shows that the optimum value of the dual problem is always an upper bound for the optimum value of the primal.20). However.

and M : B → R {+∞} is a concave function. ˜ belong to B. equivalently. m . DUALITY THEORY ˆ x is optimal solution of (5. λ) satisfy the optimality ˆ x ˆ ˆ by virtue of (5. Also. and hence by deﬁnition f0 (ˆ) = M (ˆ Also ˆ x b). . and fi . are differentiable.22) ˆ ˆ λi = 0 when fi (ˆ) < ˆi . Note that in this case x ∈ Ω(b) ˆ The next result is equivalent to Theorem 4(ii) of Section 1 if X = Rn .. . . x x (5. ˆ x b (5.21) x is feasible for (5. Then b).21) x x b) = f0 (ˆ) by (5.23) ˆ λ ≥ 0 is said to be an optimal price vector if there is x ∈ X such that (ˆ. f0 (θx + (1 − θ)˜) ≥ θf0 (x) + (1 − θ)f0 (˜) . let x ∈ Ω(b). b) ˆ Proof: Let x ∈ Ω(ˆ so that λ (f (x) − ˆ ≤ 0. ♦ We now proceed to a much more detailed investigation. ˆ 59 (5. Proof: Let b. x hence (θx + (1 − θ)˜) ∈ Ω(θb + (1 − θ)˜ x b) and therefore. then x is an optimal solution to x ˆ ˆ ˆ the primal. ˆ ˆ m(λ) = f0 (ˆ) − λ (f (ˆ) − ˆ x x b) ˆ .18) with λ = λ. λ (f (ˆ) − ˆ = 0.23) x so that x is optimal for the primal. f0 (ˆ) = M (b) x ˆ so that from Weak Duality λ is optimal for the dual.2. fi (ˆ) ≤ ˆi for i = 1.22).17). Then (θx + (1 − θ)˜) ∈ X b ˜ b).5. λ) satisfy the optimality conditions. and M (ˆ = m∗ . x since X is convex. 0 ≤ i ≤ m. . b) ˆ f0 (x) ≤ f0 (x) − λ (f (x) − ˆ b) ˆ (f (x) − ˆ ∈ X} ≤ sup{f0 (x) − λ b)|x ˆ = f0 (ˆ) − λ (f (ˆ) − ˆ by (5. and fi (θx + (1 − θ)˜) ≤ θfi (x) + (1 − θ)fi (˜) x x since fi is convex. since f0 is concave. x ∈ Ω(˜ let 0 ≤ θ ≤ 1. B is convex. x b x b) (5.24) . λ is an optimal solution to the dual. Theorem 1: (Sufﬁciency) If (ˆ. i.e. condition. Lemma 3: B is a convex subset of Rm . so that b fi (θx + (1 − θ)˜) ≤ θb + (1 − θ)˜ .

A vector λ ∈ Rn is said to be a supergradient (subgradient) of g at x ∈ X if ˆ g(x) ≤ g(ˆ) + λ (x − x) for x ∈ X.24) holds for all x ∈ Ω(b) and x ∈ Ω(˜ it follows that ˜ b) M (θb + (1 − θ)ˆ ≥ sup {f0 (θx + (1 − θ)˜)|x ∈ Ω(b). ♦ Deﬁnition: Let X ⊂ Rn and let g : X → R {∞. b ˆ b M is not stable at ˆ b .) M (b) M (b) M (ˆ b) b) .3. M is stable at ˆ if M does not increase inﬁnitely steeply in a neighborhood of ˆ See b b.) A more geometric way of thinking about subgradients is the following. Deﬁne the subset A ⊂ R1+m by . b M (b) - ˆ b M is stable at ˆ b M (ˆ + λ (b − ˆ b) b) - . Deﬁnition: The function M : B → R number K such that {∞} is said to be stable at ˆ ∈ B if there exists a real b M (b) ≤ M (ˆ + K|b − ˆ for b ∈ B . x ˆ (g(x) ≥ g(ˆ) + λ (x − x) for x ∈ X.60 CHAPTER 5. b) b| (In words. Figure 5. x ∈ Ω(˜ b) x ˜ b)} ≥ sup{f0 (x)|x ∈ Ω(b)} + (1 − θ) sup {f0 (˜)|˜ ∈ Ω(˜ x x b)} ˜ = θM (b) + (1 − θ)M (b).) x ˆ (See Figure 5-3. −∞}. NONLINEAR PROGRAMMING Since (5.∈M (ˆ b B .3: Illustration of supergradient of stability. λ is a supergradient at ˆ b Figure 5. - ˆ b - - b .

. λ1 . ˜ ≥ ˆ and r ≤ M (ˆ then ˜ ∈ B. . b) b| ♦ The next two results give important alternative interpretations of supergradients. . b) ˆ b b) ˆ x ˆ ˆ b so that λ (f (ˆ) − ˆ ≥ 0. Then λ is a supergradient of M at ˆ iff λ is an ˆ b ˆ satisfy the optimality conditions. b) ((λ1 /λ0 . . . (M (ˆ ˆ then λ0 ≥ 0. ˆ then Let λ be a supergradient at b. b)|b ∈ B. . −λm ) deﬁnes a non-vertical hyperplane b supporting A at (M (ˆ ˆ (ii) if (λ0 . . Next let x ∈ X. (λm /λ0 )) is a supergradient of M at ˆ b. . ˆ ˆ Lemma 5: Suppose that x is optimal for (5. λ ≥ 0 and ˆ By Exercise 2. ˆ x b. . . Deﬁnition: A vector (λ0 .) The supporting (In words. Then x b) x (f0 (x). ˜ b). . x b) From the Greek “hypo” meaning below or under. λm ) is said to be the normal to a hyperplane supporting A at a point(ˆ. −λ1 .2. This neologism contrasts with the epigraph of a function which is the set lying above the graph of the function. futhermore. . b) ∈ A .17). hence again by Exercise 3 ˆ M (ˆ − λ ˆ ≥ f0 (x) − λ f (x) . x M (ˆ − λ ˆ ≥ M (ˆ − λ f (ˆ) . . i=1 (5. . r b) ˆ ∈ B. b. . We call A the hypograph1 of M . and r ≤ M (b)} .4. and M (ˆ < ∞.25) λi bi }. Lemma 4: (Gale [1967]) M is stable at ˆ iff M has a supergradient at ˆ Proof: (Sufﬁciency only) b b. f (ˆ) = M (ˆ x ∈ X. 1 . ˆ if r b) m m λ0 r + ˆ i=1 b λiˆi ≥ λ0 r + λi bi for all (r. A lies below the hyperplane π = {(r. λm ) is a Exercise 3: Assume that b b) supergradient of M at ˆ then λ ≥ 0. optimal price vector. λ) x ˆ Proof: By hypothesis. The reader who is familiar with the Separation Theorem of convex sets should be able to construct a proof for the second part based on Figure 5. Exercise 2: Show that if ˆ ∈ B. DUALITY THEORY A = {(r. −λ1 . giving (5. . f (x)) ∈ A. b) ˆ b ˆ Since f0 (ˆ) = M (ˆ and λ (f (ˆ) − ˆ = 0. if the hyperplane is non-vertical then b). 61 Thus A is the set lying ”below” the graph of M . ˆ f (ˆ)) ∈ A and by Exercise 3. b)|λ0 r+ λi bi = λ0 r + ˆ ˆ hyperplane is said to be non-vertical if λ0 = 0. . Show that (i) if λ = (λ1 .23). we can rewrite the inequality above as x b). ˜ ∈ A. λi ≥ 0 for 1 ≤ i ≤ m. . (M (b). M (b) ≤ M (ˆ + λ (b − ˆ b) b) ≤ M (ˆ + |λ||b − ˆ . But then λ (ˆ − f (ˆ)) = 0. b b). Since M is concave it follows immediately that A is convex (in fact these are equivalent statements). −λm ) deﬁnes a hyperplane supporting A at b).5. and (1. . See Figure 5. b b b. . or see the Appendix at the end of this Chapter. and then (ˆ.4. M (˜ and (˜. We will prove only one part of the next crucial result. b). and f (ˆ) ≤ ˆ Let λ be a supergradient of M at ˆ x b).

A b ˆ b No non-vertical hyperplane supporting A at (M (ˆ ˆ b).e.21) by (5. b) M (b) π ˆ M (ˆ b) (λ0 .22). and (5. by (5. Then λ is a supergradient of M at ˆ iff λ is an b b) b ˆ optimal solution of the dual (5. A ˆ b b π is a non-vertical hyperplane supporting A at (M (ˆ ˆ ˆ b). suppose x ∈ X. ♦ ˆ ˆ Lemma 6: Suppose that ˆ ∈ B. . λ ≥ 0 satisfy (5. . Then λ ˆ f0 (x) ≤ f0 (x) λ (f (x) − b) ˆ = f0 (x) − λ (f (x) − ˆ + λ (b − ˆ b) ˆ b) ˆ (f (ˆ) − ˆ + λ (b − ˆ ˆ ≤ f0 (ˆ) − λ x x b) b) ˆ (b − ˆ = f0 (ˆ) + λ x b) = M (ˆ + λ (b − ˆ . ˆ ˆ (f (x) − b) ≤ 0 so that x ∈ X. . f (x) ≤ b. b) Figure 5. and M (ˆ < ∞.62 M (b) CHAPTER 5. λ) satisfy the optimality conditions. ˆ be a supergradient of M at ˆ Let x ∈ X. (5. λm ) . x ˆ ˆ Conversely. i. By Exercises 2 and 3 Proof: Let λ b.23) . b) ˆ b) ˆ so that λ is a supergradient of M at ˆ b. x x b) b) so that (5. ˆ ˆ f0 (ˆ) + λ (f (ˆ) − ˆ ≥ f0 (x) − λ (f (x) − ˆ .21).19) and m(λ) = M (ˆ b). . b) ˆ b) Hence M (b) = sup{f0 (x)|x ∈ Ω(b)} ≤ M (ˆ + λ (b − ˆ .21) holds.. Let x ∈ Ω(b). It follows that (ˆ.23). NONLINEAR PROGRAMMING M (ˆ b) .4: Hypograph and supporting hyperplane.

(Hint: See Theorem 5 of 4. b) b) so that ˆ ˆ M (ˆ ≥ sup{f0 (x) − λ (f (x) − ˆ ∈ X} = m(λ) . by Lemmas 4. λ) satisfy the (iii) if λ ˆ x ˆ optimality conditions of (5. ♦ We can now summarize our results as follows.2 Interpretation and extensions.3 imply stability. b) b)|x ˆ ˆ By weak duality (Lemma 2) it follows that M (ˆ = m(λ) and λ is optimal for (5. then λ (f (x) − b) ≤ 0. b However. in particular if there exists x ∈ X such that fi (x) < ˆi for b b 1 ≤ i ≤ m. Hence. Proof: (i) follows from Lemmas 4.23). (5.21).6.22). and (5. the various conditions of Section 1. ˆ is any optimal solution for the dual.21). Thus the condition of stability of M at ˆ plays a similar role to the constraint qualiﬁcation.3. and M is stable at ˆ Then b b) b. Theorem 2: (Duality) Suppose ˆ ∈ B. M (b) = sup{f0 (x)|x ∈ Ω(b)} ≤ M (ˆ + λ (b − ˆ . 0 ≤ i ≤ m. and m(λ) = M (ˆ Then for any x ∈ X ˆ Conversely suppose λ b).2.22). ♦ ˆ Corollary 1: Under the hypothesis of Theorem 2. then M is stable at ˆ b. .8). (5. then x is optimal for the primal iff (ˆ. (ii) is implied by Lemma 6. Lemma 7: If ˆ is in the interior of B. 6 stability is equivalent to the existence of optimal dual variables.19). It is easy to see using convexity properties that. Exercise 4: Prove Corollary 1. and (5. ˆ b. ˆ ˆ (i) there exists an optimal solution λ for the dual. ˆ M (ˆ ≥ f0 (x) − λ (f (x) − ˆ .2. b) ˆ ≥ 0. whereas the “only if” part of (iii) follows from Lemma 5. if X = Rn and fi . whereas CQ is only a sufﬁcient condition.) 5. b) b) ˆ and if moreover f (x) ≤ b.2. so that ˆ M (ˆ ≥ f0 (x) − λ (f (x) − ˆ + λ (f (x) − b) b) b) ˆ ˆ ˆb = f0 (x) − λ b + λ ˆ for x ∈ Ω(b) . then the optimality conditions (5. are differentiable. b) ˆ b) 63 ˆ so that λ is a supergradient.5. Here if X = R i we give one sufﬁcient condition which implies stability for the general case.23) are equivalent to the Kuhn-Tucker condition (5. n and the f are differentiable. The “if” part of (iii) follows from Theorem 1. if λ is an optimal solution to the dual then (∂M + /∂bi )(ˆ ≤ λi ≤ (∂M − /∂bi )(ˆ b) ˆ b). ˆ ˆ (ii) λ is optimal for the dual iff λ is a supergradient of M at ˆ b. and m(λ) = M (ˆ b). DUALITY THEORY ˆ M (ˆ − λ ˆ ≥ f0 (x) − λ f (x) b) ˆ b or ˆ M (ˆ ≥ f0 (x) − λ (f (x) − ˆ . M (ˆ < ∞. In other words if CQ holds at x then M is stable at ˆ In particular.

4. (5.18).26) . and only depends on the fact that M is concave. Much of duality theory can be given an economic interpretation similar to that in Section 4. and ˆ is the interior of B.2.17) is the short-term decision problem faced by the ﬁrm. equivalently. esis on X and fi . and comparing with Figure 5. A ˆ b b Figure 5. see Figure 5. A much more promising development has recently taken place. If for a price system λ. b). if we interpret (∂M + /∂bi )(ˆ b) b) as the marginal revenue of the ith resource.17) also is an optimal solution for (5. 1 ≤ i ≤ m.5: If M is not concave there may be no supporting hyperplane at (M (ˆ ˆ b). (5. b as the vector of current resource supplies.64 CHAPTER 5. M (ˆ < ∞ we can see from Theorem 2 and its Corollary 1 that there exists b b) an equilibrium price system iff (∂M + /∂bi )(ˆ < ∞. It is possible to obtain the duality theorem under conditions slightly weaker than convexity but since these conditions are not easily veriﬁable we do not pursue this direction any further (see Luenberger [1968]). Thus. The primal problem (5. X as constraints due to physical or long-term limitations.4. For details see the b) b Appendix. The various convexity conditions are generalizations of the economic hypothesis of non-increasing returns-to-scale. .26): Maximize f0 (x) − F (f (x) − ˆ b) subject to x ∈ X . Referring to Figure 5. . if its hypograph A is not convex. and ﬁnally f (x) the amount of these resources used up at activity levels x.18) we would then have more general problem of the form (5. M (ˆ < ∞ without loss of generality. if the current ˆ resources can be bought or sold at prices λ = (λ1 . b). b) ˆ eral than hyperplanes. f0 (x) the corresponding revenue. ˆ then we can interpret λ as a system of equilibrium prices just as in 4. These ideas are developed in (Gale [1967]). Next. NONLINEAR PROGRAMMING The proof rests on the Separation Theorem for convex sets.18). λm ) . we can think of x as the vector of n activity levels. Instead of (5. Assuming the realistic condition ˆ ∈ B.3 or Figure 5.5 it is evident that if M is not concave or. . there may be no hyperplane supporting A at (M (ˆ ˆ This is the reason why duality theory requires the often restrictive convexity hypothb). The basic idea involved is to consider supporting A at (M (ˆ ˆ by (non-vertical) surfaces π more genb). M (b) M (ˆ b) .6. we can say that equilibrium prices exist iff marginal productivities of every (variable) resource is ﬁnite. . the ﬁrm faces the decision problem ˆ an optimal solution of (5.

6) is the graph of the function b → M (ˆ − ˆ b) F (b − ˆ Usually F is chosen from a class of functions φ parameterized by µ = (µ1 . b) ≥ 0 for b ≥ 0. presented in (Frank [1969]). ˜ b) whenever b ≥ ˜ A relatively unnoticed.1 see (Geoffrion [1970a]) and for a mathematically more elegant treatment see (Rockafellar [1970]).22) and those optimal solutions of (5. In particular ˆ from Theorem 2 (iii).26): Maximize f0 (x) − φ(µ. . if we have an optimal dual solution λ then the optimal primal solutions are ˆ which also satisfy the feasibility condition (5. b)|x then the dual problem is Minimize ψ(µ) subject to µ ≥ 0 . f (x) − ˆ For b b). M (b) π ˆ (5. but quite interesting development along these lines is b. f (x) − ˆ ∈ X} .5. DUALITY THEORY 65 where F : Rm → R is chosen so that π (in Figure 5. . Parts (i) and (iii) of Theorem 2 make duality theory attractive for computation purposes. (Greenberg and Pierskalla [1970]). The following references are pertinent: (Gould [1969]). ·) then the resources f (x) − ˆ can be bought (or sold) for the amount φ(µ.2. in analogy with (5. . For more details concerning the topics of 2. f (x) − ˆ b) subject to x ∈ X . no such limitation on φ is necessary. Also see (Arrow and Hurwicz [1960]). Decentralized resource allocation. (Banerjee [1971]). 5.18) for λ = λ . of course. b).6: The surface π supports A at (M (ˆ ˆ ˆ b). If we let ψ(µ) =sup{f0 (x) − φ(µ. µk ) ≥ 0.27) would be that if the prevailing (non-uniform) price system is φ(µ.19). b). A ˆ b b Figure 5.27) instead of (5. b) ≥ φ(µ. and φ(µ. such an interpretation to make sense we should have φ(µ. Then for each ﬁxed µ ≥ 0 we have (5. The economic interpretation of (5.27) M (ˆ b) .2.3 Applications. For non-economic applications. .

This is useful because generally speaking (5.18) has fewer constraints. and the decision variable of the ith sub-system is a vector xi ∈ Rni .29) are decentralized whereas in (5. .18) is easier to solve than (5.28) whereas (5. .17) since (5.e. are strictly concave.66 CHAPTER 5. .29) involves fewer constraints. (5. . in fact. + f0 (xk ) where f0 : Rni → R are concave functions. NONLINEAR PROGRAMMING the “complementary slackness” condition (5.19) is k i Maximize f0 (xi ) − λ f i (xi ) − λ ( i=1 f i (xi ) − ˆ b) subject to xi ∈ X i . which decomposes into k separate problems: i Maximize f0 (xi ) − λ f i (xi ) i ∈X . For simplicity suppose that the i f0 .29) has an optimal solution for every λ ≥ 0. + f k (xk ) ≤ ˆ where f i : Rni → Rm are convex functions and ˆ ∈ Rm b b is the vector of available common resources. . The system is made up of k sub-systems (divisions). subject to x i k i If we let mi (λ) = sup{f0 (xi ) − λ f i (xi )|xi ∈ X i }. for each λ ≥ 0 there is a unique optimal solution of (5. The sub-system has individual constraints of the form xi ∈ X i where xi is a convex set. λ ≥ 0.29). say xi (λ). and also suppose that (5. the sub-systems share some resources in common and this limitation is expressed as f 1 (x1 ) + . if k is very large it may be practically impossible to solve (5. . but perhaps more importantly the decision problems in (5. Furthermore. mi (λ) + λ ˆ then the dual problem is Minimize m(λ) . 1 ≤ i ≤ k. f i (xi ) ≤ ˆ . Thus we have the decision problem (5. 1 ≤ i ≤ k.29) b. and m(λ) = i=1 (5.30) Note that (5. (5. xk are coupled together.29) may be trivial if the dimensions of xi are small.23). Assuming that (5. we need to ﬁnd an optimal dual solution so that we can use Theorem 2(iii). 1 ≤ i ≤ k.. .29) may be much easier to solve than (5. Consider a decision problem in a large system (e.28): k Maximize i=1 k i=1 i f0 (xi ) subject to xi ∈ X i . b (5. it is the form f0 (x1 ) + . subject to λ ≥ 0 . Then by Exercise 8 of Section 1. Consider the following algorithm.28) because. 1 ≤ i ≤ k . a multi-divisional ﬁrm).28) has an optimal solution and the stability condition is satisﬁed. 1≤i≤k . .28) For λ ∈ Rm . ﬁrst of all. Suppose that the objective function of the large system 1 k i is additive.g. the problem corresponding to (5. . i.28) all the decision variables x1 .

Select λ0 ≥ 0 arbitrary. indeed. and for other decentralization schemes for solving (5. . . The principle of conservation of mass gives us equations (5. then life in the stream is seriously threatened. Step 3.. . . optimal.2. et al. Hence. Control of water quality in a stream. qi (t) = concentration of DO measured in mg/liter. We ﬁrst derive the equations which govern the evolution in time of BOD and DO in the n areas of the streams. the stream can “die. t = 1. k Compute ep = i=1 f i (xi (λp )) − ˆ If ep ≥ 0. it is enough to study the problem over a 24-hour period. The ﬂuctuations of BOD and DO will be cyclical with a period of 24 hours. t = 1. .” Therefore. [1971]). .28) and can easily be seen to be b. namely the biochemical oxygen demand (BOD) which they place on the dissolved oxygen (DO) in the stream. Since the DO in the stream is used to breakdown chemically the pollutants into harmless substances. si (t) = concentration of BOD of efﬂuent discharge in mg/liter. For an informal discussion of schemes of pollution control which derive their effectiveness from duality theory see (Solow [1971]). We divide this period into T intervals. and go to Step 2.28) see (Geoffrion [1970b]).32) . xp will converge to the optimum solution of (5. T and i = 1.32): zi (t + 1) − zi (t) = −αi zi (t) + ψi−1 zi−1 (t) vi − ψi zi (t) vi + si (t)mi (t) vi .7 is a schematic diagram of a part of a stream into which n sources (industries and municipalities) discharge polluting efﬂuents. T . The pollutants consist of various materials. Set p = p + 1 and return to Step 3. .31) and (5. .28). it is important to treat the efﬂuents before they enter the stream in order to reduce the BOD to concentration levels which can be safely absorbed by the DO in the stream. Set λp=1 according to λp+1 = i if ep ≥ 0 λp i i p p ep if ep < 0 λi − d i i where dp > 0 is chosen a priori. Set p = 0. It is a well-advertized fact that if the DO drops below a certain concentration. N. (5. and mi (t) = amount of efﬂuent discharge in liters. xk (λp )).) Figure 5. . . but for simplicity of exposition we assume that their impact on the quality of the stream is measured in terms of a single quantity. . During interval t and in area i let zi (t) = concentration of BOD measured in mg/liter. Step 2. Solve (5. DUALITY THEORY 67 Step 1. (5. . For more detail see (Arrow and Hurwicz [1960]). In this example we are concerned with ﬁnding the optimal balance between costs of waste treatment and costs of high BOD in the stream.31) s qi (t + 1) − qi (t) = βi (qi − qi (t)) + ψi−1 qi−1 (t) − ψi qi (t) vi vi +αi zi (t) − ηi vi . xp is feasible for (5. . The discussion in this section is mainly based on (Kendrick.5. . See (Dorfman and Jacoby [1970]. the quality of the stream improves with the amount of DO and decreases with increasing BOD. It can be shown that if the step sizes dp are chosen properly. .29) for λ = λp and obtain the optimal solution xp = (x1 (λp ). .

The costs associated with increased amounts of BOD and reduced amounts of DO are much more difﬁcult to quantify since the stream is used by many institutions for a variety of purposes (e. qi (1).68 CHAPTER 5. zN qN (1 (1 − πi )si − πi+1 )si+1 (1 − πN )sN si si−a si si+1 sN Figure 5. Finally. We further assume that f is convex. (5. . agricultural. . industrial.31) is replaced by zi (t + 1) − zi (t) = −αi zi (t) + ψi zi−1 vi − ψi zi (t) vi + (1−πi (t))si (t)mi (t) vi . mi (t)) where the function must be monotonically increasing in all of its arguments.. recreational). the cost in period t will be fi (πi (t). zi−1 qi−1 . instead of attempting to quantify these costs let us suppose that some minimum water quality standards are set. si (t).7: Schematic of stream with efﬂuent discharges. Then (5. Hence. Also z0 (t). .g. . βi is the rate of generation of DO. ψi . αi . Therefore.. i = 1. q0 (t) which are the concentrations immediately upstream from area 1 are assumed known. The vi .. ηi is the DO requirement in the bottom sludge. municipal. the initial concentrations zi (1). They may vary with the time interval t. q s are parameters of the stream and are assumed known.. Finally.. This decay occurs by combination of BOD and DO. vi = volume of water in area i measured in liters. Let q be the minimum acceptable DO concentration and let z be the maximum permissible BOD concentration. Here. The cost of waste treatment can be readily identiﬁed. N are assumed known. ηi .. ψi = volume of water which ﬂows from area i to are i + 1 in each period measured in liters.33) We now turn to the costs associated with waste treatment and pollution. Now suppose that the waste treatment facility in area i removes in interval t a fraction πi (t) of the concentration si (t) of BOD. . . αi is the rate of decay of BOD per interval. In period t the ith facility treats mi (t) liters of efﬂuent with a BOD concentration si (t) mg/liter of which the facility removes a fraction πi (t). and the disutility caused by a decrease in the water quality varies with the user. Then we face the ¯ . The increase in DO is due to various natural oxygen-producing biochemical reactions in the stream and the increase is proportional to (q s − qi ) where q s is the saturation level of DO in the stream. NONLINEAR PROGRAMMING direction of ﬂow 0 z0 q0 (1 − π1 )si given 1 z1 q1 1 − πi−1 i−1 i zi qi i+1 zi+1 qi+1 N N +1 .

i = 1. it does not make sense to enforce legally a minimum standard qi (t) ≥ q. t = 1. . i = 1. . furthermore. Then assuming that the agency is required to maintain the standards (q. .5. where the 2N T -dimensional vector w has its components equal to −q or z in the obvious manner. . . .e. . t = 1.34)). . . t = 1. N . and −qi (t) ≤ −q . . and let w = (w(1). N . On the other hand. The question we now pose is whether there exist tax rates such that if each individual polluter minimizes its own total cost (i. N .33) for w and obtain w = b + Ar . . . . . .36) subject to b + Ar ≤ w . . . T. T. and r is the NT-dimensional vector with components (1 − πi (t))si (t)mi (t).35) we can rewrite (5. . si (t). Suppose that all the treatment facilities are in the control of a single public agency. It should be clear from the duality theory that the answer is in the afﬁrmative.34) subject to (5. T . DUALITY THEORY following NP: N T 69 Maximize − i=1 t=1 fi (πi (t). .2. . (5. let w(t) = (w1 (t). cost of waste treatment + tax on remaining pollutants). i = 1. zi (t) ≤ z on every polluter since the ¯ pollution levels in the ith area depend upon the pollution levels on all the other areas lying upstream. . of the problem: Maximize − i t fi (πi (t).35) where the matrix A and the vector b depend upon the known parameters and initial conditions. zi (t) ≤ z . z ).34) and arrive at an optimal solution. . mi (t)) (5. i = 1. To see this let wi (t) = (zi (t). −qi (t)) .36) and. . . Note that the coefﬁcients of the matrix must be non-negative because an increase in any component of r cannot decrease the BOD levels and cannot increase the DO levels.32). . . then the resulting water quality will be acceptable and. w(t)). . N. . . Using (5. . t = 1. . N . . . . . . t = 1. T. i = 1. T . then the individual polluters may not (and usually do not) have any incentive to cooperate among themselves to achieve these standards. . . it may be economically and politically acceptable to tax individual polluters in proportion to the amount of pollutants discharged by the individual. . Furthermore.e. the optimal values of (5. ¯ ¯ ∗ ≥ 0. . . . ¯ 0 ≤ πi (t) ≤ 1 . and we write the components of p∗ as p∗ (t) to match i .36) and (5. . . . (5. si (t). . . .33).32) and (5. . wN (t)). i = 1. and an optimal solution By the duality theorem there exists a 2N T -dimensional vector λ ∗ πi (t). si (t). furthermore. (5.34) as follows: Maximize − i t fi (πi (t). If we let p∗ = A λ∗ ≥ 0. Then we can solve (5. mi (t)) − λ∗ (b + Ar − w) subject to 0 ≤ πi (t) ≤ 1. N . . T.. will be an optimal solution of (5. .37) ∗ such that {πi (t)} is also an optimal solution of (5. . t = 1. z ) and it does this at a minimum cost it will ¯ solve the NP (5. . . But if ¯ there is no such centralized agency. .. ¯ 0 ≤ πi (t) ≤ 1 . Let the minimum cost be m(q. mi (t)) (5. the resulting amount of waste treatment is carried out at the minimum expenditure of resources (i.37) are equal. .

y ∗ . . (5. .39) where x ∈ Rn is the decision variable .3 Quadratic Programming An important special case of NP is the quadratic programming (QP) problem: Maximize c x − 1 x P x 2 subject to Ax ≤ b. 5. x ≥ 0 y ≥ 0.39) iff there is a solution (x∗ . . . i Before we leave this example let us note that the optimum dual variable or shadow price λ∗ plays an important role in a larger framework. λ ≥ 0. Now suppose it is proposed to change the standard in the ith area during period t to q + ∆qi (t) and z + ∆zi (t).2).2 that f0 : x → c x − 1/2 x P x is a concave function.37) is equivalent to the set of N T problems: Maximize − fi (πi (t).2. (λ∗ ) (Ax∗ − b) = 0 .39) iff there exist λ∗ ∈ Rm . . We noted earlier that the quality standard (q. NONLINEAR PROGRAMMING with the components (1 − πi (t))si (t)mi (t) of r we can see that (5. mi (t)) − p∗ (t)(1 − πi (t))si (t)mi (t) i 0 ≤ πi (t) ≤ 1 . If the corresponding components of λ∗ are λq∗ (t) and λz∗ (t). This estimate can now serve as a basis in making a beneﬁts/cost i i analysis of the proposed new standard. λ∗ . N . (µ∗ ) x∗ = 0 . and (5.42) by Phase I of the Simplex algorithm (see 4. .λ y = 0 . T . hence the necessity of these conditions follows from Theorem 2 of 1. . (5.41) and (5. such that Ax∗ ≤ b. (5.41).38) Thus. b ∈ Rm are ﬁxed.42). si (t). CQ is satisﬁed. µ∗ ≥ 0 .43) Suppose we try to solve (5. t = 1. Then we must apply Phase II to the LP: m n Maximize − i=1 zi − j=1 ξj .70 CHAPTER 5. (5.40) Proof: By Lemma 3 of 1. A is a ﬁxed m × n matrix and P = P is a ﬁxed positive semi-deﬁnite matrix.3.41) (5. x∗ ≥ 0 c − P x∗ = A λ∗ − µ∗ . µ ≥ 0 . λ∗ ≥ 0.42) (5. x ≥ 0 . c ∈ Rn . . µ x = 0 . p∗ (t) is optimum tax per mg of BOD in area i during period t. µ∗ ∈ Rn . so that the sufﬁciency of these conditions follows from Theorem 4 of 1.43): Ax + Im Y = b −P x − A λ + In µ = −c . since P is positive semi-deﬁnite it follows from Exercise 6 of Section 1.40) we can see that x∗ is optimal for (5. i = 1. Theorem 1: A vector x∗ ∈ Rn is optimal for (5.3. ♦ From (5. z ) ¯ was somewhat arbitrary. µ∗ ) to (5. On the other hand.2. ¯ i i then the change in the minimum cost necessary to achieve the new standard will be approximately λq∗ (t)∆qi (t) + λz∗ (t)∆zi (t). (5.

µ) are non-zero.43). Furthermore.42). and (5. The algorithm will stop in a ﬁnite number of steps at an ˆ optimal basic feasible solution (ˆ. λ. .3. y . in order to obtain a solution of (5. Let Ω ⊂ Rn denote the set of feasible solutions. (5. i = 1.44) starting with a basic feasible solution z = b. If z = 0 or ξ ˆ ˆ solution to (5.45). y . ξ = 0 is ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ an optimal solution of (5. λ. do not consider λi as a candidate for entry into the basis.39). λ. . But then a repetition of the proof of Lemma 1 of 4. (5. then there is an optimal basic feasible solution of (5.2 to solve (5.43).41). (5. then there is no (5. f1 (ˆ) + f1x (ˆ)h.4 Computational Method We return to the general NP (5.43) have a solution. .41). µ. and there is no feasible solution of (5. without loss of generality. However.42) have a solution then the maximum value in (5. . p. y ≥ 0. starting with the basic feasible solution z = b. µ.43) and x is an optimal solution of (5. Theorem 2: Suppose P is positive deﬁnite. .41). x x −ψ(ˆ)(h) + fi (ˆ)fix h ≤ 0 . The above algorithm is due to Wolfe [1959]. stop. (5. For x ∈ Ω deﬁne the function ψ(ˆ) : Rn → R by ˆ x ψ(ˆ)(h) = max{−f0x (ˆ)h. y . do not consider µj as a candidate for entry into the basis. m . COMPUTATIONAL METHOD subject to Ax + Im y +z =b −P x − A λ + In µ + ξ = −c x ≥ 0.41).43). ξ ≥ 0.46) .41). 0 ≤ i ≤ m.3. from (5.45) where x ∈ Rn . and (5.44) which is also a solution f (5.44). λ. ˆ Proof: Let x.42). and (5. and Polak [1970].42).5. (5. Maximize f0 (x) subject to fi (x) ≤ 0. (5. y . fi : Rn → R. 71 (5. and (5. µ be a solution of (5.) If (5. The behavior of the algorithm is summarized below.1 will x ˆ ˆ ˆ also prove this lemma. (5. z . z = 0. y . ♦ This lemma suggests that we can apply the Simplex algorithm of 4.43): If a variable xj is currently in the basis. Then x.4. . We have the following result. Step 2 of the Simplex algorithm must be modiﬁed as follows to satisfy (5. .42). that b ≥ 0 and −c ≥ 0. (5.44). ξ = −c.39). 159 ff).44). are differentiable. z ≥ 0. fm (ˆ) + fmx (ˆ)h}. . (We have assumed.41). µ) solve x ˆ ˆ ˆ ˆ ˆ ˆ x ˆ ˆ ˆ ˆ = 0. −1 ≤ hj ≤ 1 . 5. For a proof of this result as well as for a generalization of the algorithm which permits positive semi-deﬁnite P see (Cannon. If it not possible to remove the zi and ξj from the basis. µ ≥ 0.42) and (5.42). and (5. Cullum.43) we see that at most (n + m) components of (ˆ.41) and (5. If z = 0 and ξ = 0 then (ˆ. ξ) of (5. x x x x x x Consider the problem: Minimize ψ(ˆ)(h) x subject to − ψ(ˆ)(h) − f0x (ˆ)h ≤ 0 . 1 ≤ j ≤ n . λ ≥ 0.44) is 0. (5. Lemma 1: If (5.42). if a variable yi is currently in the basis. x x 1 ≤ i ≤ m . ξ = −c. λ.43).

set k = 0. . Theorem 1: Suppose that the set Ω(x0 ) = {x|x ∈ Ω. Remark: If h0 (xk ) < 0 in Step 2.46) for x = xk and obtain h0 (xk ).) . then the direction h(xk ) satisﬁes f0x (xk )h(xk ) > 0. generated by the algorithm. The performance of the algorithm is summarized below. and go to Step 4. x1 .8. For this reason h(xk ) is called a (desirable) feasible direction.) The following algorithm is due to Topkis and Veinott [1967].8: h(xk ) is a feasible direction.72 CHAPTER 5. . and has a non-empty interior. Find x0 ∈ Ω. NONLINEAR PROGRAMMING f0 (xk )f0 (x) = F0 (x∗ ) > f0 (xk ) f1 (xk ) f0 (x) = f0 (xk ) f2 = 0 . Step 4. . . Then the Kuhn-Tucker conditions are satisﬁed at x∗ .46) and let h0 (ˆ) = ψ(ˆ)(h(ˆ)) be the minimum value atx x x x tained. If h0 (xk ) = 0. Let x∗ be any limit point of the sequence x0 . . otherwise go to Step ˆ 3. µ ≥ 0 . Compute an optimum solution µ(xk ) to the one-dimensional problem. . For a proof of this result and for more efﬁcient algorithms the reader is referred to (Polak [1971]). and go to Step 2. Step 3. which is dense in Ω(x0 ). (See Figure 5. subject to (xk + µh(xk )) ∈ Ω. Solve (5. Call h(ˆ) an optimum solution of (5. set k = k + 1 and return to Step 2. and fi (xk )+ fix (xK )h(xk ) < 0. . 1 ≤ i ≤ m. . Step 2. xk . Step 1. f0 (x) ≥ f0 (x0 )} is compact. Maximize f0 (xk + µh(xk )) . f3 (xk ) xk h(xk ) f2 (xk ) f1 = 0 Ω f3 = 0 Figure 5. (Note that by Exercise 1 of 4.1 (5. h(xk ). Set xk+1 = xk + µ(xk )h(xk ).46) can be solved as an LP.5. stop.

49) λ0 r + i=1 b λiˆi ≥ θ .48) we get m 1 λ0 [θ m M (b) − M (ˆ ≤ b) − i=1 λ i bi ] = i=1 (− λi )(bi − ˆ b).5. Separation theorem for convex sets. . G be convex subsets of Rn such that the relative interiors of F. .49) we can see that m i=1 b λiˆi ≥ θ. Evidently.7 of Section 2 are based on the following extremely important theorem (see Rockafeller [1970]).47) λ0 r + i=1 m λi bi ≤ θ for (r. r ≤ M (b) − M (ˆ . b) ∈ F . Then there exists λ ∈ Rn . G are convex and F ∩ G = φ. G are convex. Also from (5. and θ such that m (5. b)} It is easy to check that F.48) m b λiˆi ≤ θ. . b) (5. . whereas from (5. . .47) implies that F ∩ G = φ. . λm ) = 0. . so that there exist (λ0 . λm ) = 0. But then from (5. for r > M (ˆ .5. and θ such that m (5. Proof of Lemma 4: Since M is stable at ˆ there exists K such that b M (b) − M (ˆ ≤ K|b − ˆ for all b ∈ B . .5 Appendix The proofs of Lemmas 4. APPENDIX 73 5. ˆ > M (ˆ b)|r b} G = {(r. b| In R1+m consider the sets F = {(r.49) can hold m only if λ0 > 0. λ0 ♦ Proof of Lemma 7: Since ˆ is in the interior of B. i=1 so that i=1 b λiˆi = θ. i=1 λ0 r + From the deﬁnition of F . r > K|b − ˆ . .48) λi bi ≥ θ for (r.50) . b)|b ∈ B. b|} G = {(r. b) b| In R1+m consider the sets F = {(r. F. . and the fact that (λ0 . λm ) = 0. λ = 0. there exists ε > 0 such that b b ∈ B whenever |b − ˆ < ε . (5. G are disjoint. b)|b ∈ Rm . it can be veriﬁed that (5. Hence. Let F. b)|b ∈ B. b) ∈ G . and θ ∈ R such that λ g ≤ θ for all g ∈ G λ f ≥ θ for all f ∈ F . . there exist (λ0 . r ≤ M (b)} . and (5.

b λ0 ♦ . .49).51) imply λ0 > 0. .50).51) we get λ0 M (ˆ + b) so that (5. . for (r. (5. b) ∈ G .74 m CHAPTER 5. and the fact that (λ0 . λm ) = 0 we can see that (5.52) implies M (b) ≤ (ˆ + b) m m i=1 b λiˆi = θ .51) From (5. . From (5. NONLINEAR PROGRAMMING λ0 r + i=1 b λiˆi ≤ θ .(5.50) and (5. (− i=1 λi )(bi − ˆi ) .

gravitational force = gx3 with g assumed constant. 6.1. and thrust force CT x3 . Thus. at rest. Speciﬁcally suppose that the equations of motion are given by (6.Chapter 6 SEQUENTIAL DECISION PROBLEMS: DISCRETE-TIME OPTIMAL CONTROL In this chapter we apply the results of the last two chapters to situations where decisions have to be made sequentially over time. that is. and in Section 2 we derive the main result.1). the 75 . x3 (0)) = (0. See Figure 6. These equations can be derived from the force equations under the assumption that there are four forces acting on the rocket. ρ(x1 ) is a friction coefﬁcient depending on atmospheric 2 density which is a function of x1 . namely: inertia = x3 x1 = x3 x2 . M ). with initial fuel of weight M . x2 (t) is the (vertical) speed at time t. The “dot” denotes differentiation with respect to t. it is desired that the rocket be at a position as high above the ground as possible.1 Examples The trajectory of a vertical sounding rocket is controlled by adjusting the rate of fuel ejection which generates the thrust force. assumed proportional to rate of fuel ejection.1) where x1 (t) is the height of the rocket from the ground at time t. A very important class of problems where such situations arise is in the control of dynamical systems. At a prescribed ﬁnal time tf . drag ¨ ˙ force = CD ρ(x1 )x2 where CD is constant. ˙ CT x3 (t) u(t) (6. x3 (t) is the weight of the rocket (= weight of remaining fuel) at time t. At time 0 we assume that (x1 (0). x2 (0). In the ﬁrst section we give two examples. The decision variable at ˙ time t is u(t). the rocket is on the ground. 0. the rate of fuel ejection. x1 (t) = x2 (t) ˙ x2 (t) = − xCD ρ(x1 (t))x2 (t) − g + ˙ 2 3 (t) x3 (t) = −u(t) .

76

CHAPTER 6. DISCRETE-TIME OPTIMAL CONTROL

decision problem can be formalized as (6.2). Maximize x1 (tf ) subject to x(t) = f (x(t), u(t)), 0 ≤ t ≤ tf ˙ x(0) = (0, 0, M ) u(t) ≥ 0, x3 (t) ≥ 0, 0 ≤ t ≤ tf ,

(6.2)

where x = (x1 , x2 , x3 ) , f : R3+1 → R3 is the right-hand side of (6.1). The constraint inequalities u(t) ≥ 0 and x3 (t) ≥ 0 are obvious physical constraints.

x3 x1 = inertia ¨

CD ϕ(x1 )x2 = drag 2

gx3 = gravitational force ˙ CR x3 = thrust Figure 6.1: Forces acting on the rocket. The decision problem (6.2) differs from those considered so far in that the decision variables, which are functions u : [0, tf ] → R, cannot be represented as vectors in a ﬁnite-dimensional space. We shall treat such problems in great generality in the succeeding chapters. For the moment we assume that for computational or practical reasons it is necessary to approximate or restrict the permissible function u(·) to be constant over the intervals [0, t1 ), [t1 , t2 ), . . . , [tN −1 , tf ), where t1 , t2 , . . . , tN −1 are ﬁxed a priori. But then if we let u(i) be the constant value of u(·) over [ti , ti+1 ), we can reformulate (6.2) as (6.3): Maximize x1 (tN )(tN = tf ) subject to x(ti+1 ) = g(i, x(ti ), u(i)), i = 0, 1, . . . , N − 1 x(t0 ) = x(0) = (0, 0, M ) u(i) ≥ 0, x3 (ti ) ≥ 0, i = 0, 1, . . . , N .

(6.3)

In (6.3) g(i, x(t1 ), u(i)) is the state of the rocket at time ti+1 when it is in state x(ti ) at time ti and u(t) ≡ u(i) for ti ≤ t < ti+1 . As another example consider a simple inventory problem where time enters discretely in a natural fashion. The Squeezme Toothpaste Company wants to plan its production and inventory schedule for the coming month. It is assumed that the demand on the ith day, 0 ≤ i ≤ 30, is d1 (i) for

6.2. MAIN RESULT

77

their orange brand and d2 (i) for their green brand. To meet unexpected demand it is necessary that the inventory stock of either brand should not fall below s > 0. If we let s(i) = (s1 (i), s2 (i)) denote the stock at the beginning of the ith day, and m(i) = (m1 (i), m2 (i)) denote the amounts manufactured on the ith day, then clearly s(i + 1) + s(i) + m(i) − d(i) , where d(i) = (d1 (i), d2 (i)) . Suppose that the initial stock is s, and the cost of storing inventory s ˆ for one day is c(s) whereas the cost of manufacturing amount m is b(m). The the cost-minimization decision problem can be formalized as (6.4):
30

Maximize
i=0

(c(s(i)) + b(m(i))) (6.4)

subject to s(i + 1) = s(i) + m(i) − d(i), 0 ≤ i ≤ 29 s(0) = s ˆ s(i) ≥ (s, s) , m(i) ≥ 0, 0 ≤ i ≤ 30 .

Before we formulate the general problem let us note that (6.3) and (6.4) are in the form of nonlinear programming problems. The reason for treating these problems separately is because of their practical importance, and because the conditions of optimality take on a special form.

6.2 Main Result
The general problem we consider is of the form (6.5).
N −1

Maximize
i=0

f0 (i, x(i), u(i))

subject to dynamics : x(i + 1) − x(i) = f (i, x(i), u(i)), i = 0, . . . , N − 1 , initial condition: q0 (x(0) ≤ 0, g0 (x(0)) = 0 , ﬁnal condition: qN (x(N )) ≤ 0, gN (x(N )) = 0 , state-space constraint: qi (x(i)) ≤ 0, i = 1, . . . , N − 1 , control constraint: hi (u(i)) ≤ 0, i = 0, . . . , N − 1 .

(6.5)

Here x(i) ∈ Rn , u(i) ∈ Rp , f0 (i, ·, ·) : Rn+p → R, f (i, ·, ·) : Rn+p → Rn , qi : Rn → Rmi , gi : Rn → R i , hi : Rp → Rsi are given differentiable functions. We follow the control theory terminology, and refer to x(i) as the state of the system at time i, and u(i) as the control or input at time i. We use the formulation mentioned in the Remark following Theorem 3 of V.1.2, and construct the Lagrangian function L by L(x(0), . . . , x(N ); u(0), . . . , u(N − 1); p(1), . . . , p(N ); λ0 , . . . , λN ; α0 , αN ; γ 0 , . . . , γ N −1 )

78
N −1 N −1

CHAPTER 6. DISCRETE-TIME OPTIMAL CONTROL
=
i=0 N i i=0 0

f0 (i, x(i), u(i)) −
i=0

(p(i + 1)) (x(i + 1) − x(i) − f (i, x(i), u(i)))+
N −1 N

(λ ) qi (x(i)) + (α ) g0 (x(0)) + (α ) gN (x(N )) +
i=0

(γ i ) hi (u(i))

.

Suppose that CQ is satisﬁed for (6.5), and x∗ (0), . . . , x∗ (N ); u∗ (0), . . . , u∗ (N − 1), is an optimal solution. Then by Theorem 2 of 5.1.2, there exist p∗ (i) in Rn for 1 ≤ i ≤ N, λi∗ ≥ 0 in Rmi for 0 ≤ i ≤ N, αi∗ in R i for i = 0, N, and γ i∗ ≥ 0 in Rsi for 0 ≤ i ≤ N − 1, such that (A) the derivative of L evaluated at these points vanishes, and (B) λi∗ qi (x∗ (i)) = 0 for 0 ≤ i ≤ N , γ i∗ hi (u∗ (i)) = 0 for 0 ≤ i ≤ N − 1 . We explore condition (A) by taking various partial derivatives. Differentiating L with respect to x(0) gives f0x (0, x∗ (0), u∗ (0)) − {−(p∗ (1)) − (p∗ (1)) [fx (0, x∗ (0), u∗ (0))] +(λ0∗ ) [q0x (x∗ (0))] + (α0∗ ) [g0x (x∗ (0))]} = 0 , or p∗ (0) − p∗ (1) = [fx (0, x∗ (0), u∗ (x))] p∗ (1) +[f0x (0, x∗ (0), u∗ (0))] − [q0x (x∗ (0))] λ0∗ , where we have deﬁned p∗ (0) = [g0x (x∗ (x))] α0∗ . Differentiating L with respect to x(i), 1 ≤ i ≤ N − 1, and re-arranging terms gives p∗ (i) − p∗ (i + 1) = [fx (i, x∗ (i), u∗ (i))] p∗ (i + 1) +[f0x (i, x∗ (i), u∗ (i))] − [qix (x∗ (i))] λi∗ . Differentiating L with respect to x(N ) gives, p∗ (N ) = −[gN x (x∗ (N ))] αN∗ − [qN x (x∗ (N ))] λN∗ . It is convenient to replace αN∗ by −αN∗ so that the equation above becomes (6.9) p∗ (N ) = [gN x (x∗ (N ))] αN∗ − [qN x (x∗ (N ))] λN∗ . Differentiating L with respect to u(i), 0 ≤ i ≤ N − 1 gives [f0u (i, x∗ (i), u∗ (i))] + [fu (i, x∗ (i), u∗ (i))] p∗ (i + l) − [hiu (u∗ (i))] γ i∗ = 0 . We summarize our results in a convenient form in Table 6.1 Remark 1: Considerable elegance and mnemonic simpliﬁcation is achieved if we deﬁne the Hamiltonian function H by (6.10) (6.9) (6.8) (6.7)

(6.6)

. . . . u(i)) Table 6. . u∗ (i))] + [fu (i. x∗ (i). . λ0∗ . x∗ (i). p∗ (i1 ) = [hiu (u∗ (i))] γ i∗ adjoint equations: i = 0. α0∗ . . . x(i). . N − 1 x(i + 1) − x(i) = f (i. g0 (x∗ (0)) = 0 ﬁnal conditions: qN (x∗ (N )) ≤ 0. N − 1 p∗ (i) − p∗ (i + 1) = [fx (i. αN∗ . . . . . . u∗ (i)] p∗ (i + 1) +[f0x (i. . u∗ (i)] − [qix (x∗ (i)] γ i∗ transversality conditions: p∗ (0) = [g0x (x∗ (0))] α0∗ λ0∗ ≥ 0. . . N − 1 hi (u∗ (i)) ≤ 0 [f0u (i. λN∗ . . x∗ (N ). . . . x(i). . u∗ (0). . (λi∗ ) qi (x∗ (i)) = 0 γ i∗ ≥ 0 (γ i∗ ) hi (u∗ (i) = 0 79 . . (λ0∗ ) q0 (x∗ (0)) = 0 λN∗ ≥ 0. such that f0 (i.1: initial condition: q0 (x∗ (0)) ≤ 0. γ N −1∗ . x∗ (i). .2. . u(i)) subject to the constraints below dynamics: i = 0. γ 0∗ . . p∗ (N ) = [gN x (x∗ (N ))] αN∗ − [qN x (x∗ (N ))] λN∗ (λN∗ ) qN (x∗ (N )) = 0 λi∗ ≥ 0. N − 1 qi (x∗ (i)) ≤ 0 control constraint: i = 0. gN (x∗ (N )) = 0 state space constraint: i = 1.6. MAIN RESULT Suppose x∗ (0). x∗ (i)u∗ (i))] . . . u∗ (N − 1) maximizes N −1 i=0 then there exist p∗ (N ). . . .

Furthermore. p∗ (i + 1)). which describe surfaces in Rn . Remark 4: The conditions (6. Furthermore. (6. For this reason the result is sometimes called the maximum principle. x∗ (i).7). x∗ (i). x∗ (i). Remark 2: If we linearize the dynamic equations about the optimal solution we obtain δx(i + 1) − δx(i) = [fx (i. x∗ (i). x∗ (i). subject to hi (u) ≤ 0 . . u∗ (i). (6.8) become (6. p) = f0 (i. DISCRETE-TIME OPTIMAL CONTROL H(i. u∗ (i))]δx(i) + [fu (i. we call (6. x∗ (i).13). The dynamic equations then become x∗ (i + 1) − x∗ (i) = [Hp (i. qN ≡ 0. x∗ . ·) are concave and the remaining function in (6. (i). x∗ (i. (6. which has for it adjoint the system r(i) − r(i + 1) = [fx (i.5). x. x∗ (i).13) that u∗ (i) is an optimal solution of Maximize H(i.9) are called transversality conditions for the following reason. u∗ (i).12). i. ·. and the adjoint equations (6.13) (6. Conditions (6. x. so that the initial and ﬁnal conditions read g0 (x(0)) = 0.9) become respectively p∗ (0) = [g0x (x∗ (0))] α0∗ .6).7). u∗ (i))]z(i) .5) are linear. p∗ (i + 1))] . u) + p f (i. 0 ≤ i ≤ N − 1 . and the p∗ (i) are called adjoint variables. but note that these 2n boundary conditions are mixed. (6. u∗ (i))]δu(i) . x.6) and (6.8) is (6. u∗ (i))] r(i + 1) .. in this case we see from (6. we note that in this case the initial and ﬁnal conditions specify ( 0 + n ) conditions whereas the transversality conditions specify (n − 0 ) + (n − n ) conditions. (6. u.e. Suppose q0 ≡ 0. Remark 3: If the f0 (i.6). whereas (6.80 CHAPTER 6.1 are also sufﬁcient.10) becomes [hiu (u∗ (i))] γ i∗ = [Hu (i. and the necessary conditions of Table 6. p∗ (N ) = [gN x (x(N ))] αN∗ which means that p∗ (0) and p∗ (N ) are respectively orthogonal or transversal to the initial and ﬁnal surfaces. whose homogeneous part is z(i + 1) − z(i) = [fx (i. 0≤i≤N −1. gN (x(N )) = 0.12) Since the homogeneous part of the linear difference equations (6.8) the adjoint equations. p∗ (i + 1))] . u) . 0≤i≤N −1 . Thus. p∗ (i + 1))] − [qix (x∗ (i))] λi∗ . we have a total of 2n boundary conditions for the 2n-dimensional system of difference equations (6. u.11) p∗ (i) − p∗ (i + 1) = [Hx (i. then CQ is satisﬁed. u∗ (i). (6. some of them refer to the initial time 0 and the rest refer to the ﬁnal time. u∗ (i).

1 2 N −1 N −1 81 x(i) Qx(i) + where x(i) ∈ Rn . x(N ) = x(N ) . 0 ≤ i ≤ N − 1 . Exercise 2: Show that the minimal fuel problem.6. show that the optimal solution is unique and can be obtained by solving a 2n-dimensional linear difference equation with mixed boundary conditions.2. Here x(0). Q = Q is positive semi-deﬁnite. 0 ≤ i ≤ N − 1 x(0) = x(0). j=1 subject to x(i + 1) − x(i) = Ax(i) + Bu(i). 1 ≤ j ≤ p. A and B are as ˆ ˆ in Exercise 1. 0 ≤ i ≤ N − 1 x(0) = x(0). x(N ) are ﬁxed. 0 ≤ i ≤ N − 1 can be transformed into a linear programming problem. |u(i))j | ≤ 1. ˆ and P = P is positive deﬁnite. Maximize 1 u(i) P u(i) 2 i=0 i=0 subject to x(i + 1) − x(i) = Ax(i) + Bu(i). x(0) is ﬁxed. ˆ u(i) ∈ Rp . A and B are constant matrices.   N −1 i=0 Minimize  P |(u(i))j | . MAIN RESULT Exercise 1: For the regulator problem. ˆ ˆ u(i) ∈ Rp . .

82 CHAPTER 6. DISCRETE-TIME OPTIMAL CONTROL .

˙ where x(t) ∈ Rn and u(t) ∈ Rp are respectively the state and control of the system at time t. x0 ∈ Rn be ﬁxed and let tf ≥ t0 be a ﬁxed time. ∞) → Ω will be called an admissible control. we assume that they are piecewise continuous functions. ˙ (7. In Section 2 we study more general boundary conditions.Chapter 7 SEQUENTIAL DECISION PROBLEMS: CONTINUOUS-TIME OPTIMAL CONTROL OF LINEAR SYSTEMS We will investigate decision problems similar to those studied in the last chapter with one (mathematically) crucial difference. and to be piecewise continuous. We are concerned with the 83 . 7. In Section 1 we present the general linear problem and study the case where the initial and ﬁnal conditions are particularly simple. Deﬁnition: A piecewise continuous function u : [t0 . A choice of control has to be made at each instant of time t where t varies continuously over a ﬁnite interval. The general nonlinear case is deferred to the next chapter. U denotes the set of all admissible controls. x(t). To understand the main ideas and techniques of analysis it will prove proﬁtable to study the linear case ﬁrst.1 The Linear Optimal Control Problem We consider a dynamical system governed by the linear differential equation (7.1) Here A(·) and B(·) are n × n.1): x(t) = A(t)x(t) + B(t)u(t). The evolution in time of the state of the systems to be controlled is governed by a differential equation of the form: x(t) = f (t. t ≥ t0 . The control u(·) is constrained to take values in a ﬁxed set Ω ⊂ Rp .and n × p-matrix valued functions of time. u(t)) . Let c ∈ Rn .

tf ] → Rp . The next result is well-known. z. t1 )z + t2 t1 Φ(t2 .1). t) ≡ In . τ ). u)|u ∈ U} . t1 . z.2). (ii) Let K(t2 .1. t1 . ﬁnal condition: x(tf ) ∈ Rn . be the transition-matrix function of the homogeneous part of (7. We say that c is the outward normal to a hyperplane supporting K at x∗ if c = 0. Thus. t1 . t0 ≤ t ≤ tf .2) Deﬁnition: (i) For any piecewise continuous function u(·) : [t0 . τ )B(τ )u(τ )dτ . CONTINUOUS-TIME LINEAR OPTIMAL CONTROL Maximize c x(tf ). (7. t1 . but then . z) is the set of states reachable at time t2 starting at time t1 in state z and using admissible controls. if a time t1 it is in state z. u∗ ). i.) Deﬁnition: Let K ⊂ Rn . t1 . x0 ). u) = Φ(t2 . Let u∗ (·) ∈ U and let x∗ (t) = φ(t.1) at time t2 . t0 . τ ) . Exercise 1: (i) Assuming that Ω is convex. Lemma 2: Suppose c = 0. and the control u(·) is applied. Φ satisﬁes the differential equation ∂Φ ∂t (t. for any z ∈ Rn . We call K the reachable set. z) = {φ(t2 . control constraint: u(·) ∈ U . t1 . and (ii) c is the outward normal to a hyperplane supporting K at x∗ . K(t2 . (See Desoer [1970].) Lemma 1: φ(t2 . ˙ initial condition: x(t0 ) = x0 . and the boundary condition Φ(t. (ii) Assuming that U is convex show that K(t2 . (See Figure 7. z) is a convex set. u) denote the state of (7.84 decision problem (7. z) is convex even if Ω is not convex (see Neustadt [1963]). x0 .) Proof: Clearly (i) is implied by (ii) because if x∗ (tf ) is in the interior of K there is δ > 0 such that (x∗ (tf ) + δc) ∈ K.e. show that U is a convex set. Then u∗ is an optimal solution of (2) iff (i) x∗ (tf ) is on the boundary of K = K(tf . τ ) = A(t)Φ(t. (It is a deep result that K(t2 . provided we include in U any measurable function u : [t0 . t1 . and any t0 ≤ t1 ≤ t2 ≤ tf let φ(t2 . The next result gives a geometric characterization of the optimal solutions of (2). CHAPTER 7. Deﬁnition: Let Φ(t. t0 . ∞) → Ω. and c x∗ ≥ c x for all x ∈ K . z. and let x∗ ∈ K. t0 ≤ τ ≤ t ≤ tf . subject to dynamics: x(t) = A(t)x(t) + B(t)u(t) ..

τ )B(τ )u(τ )dτ (7.7. τ )B(τ )u (τ )dτ t ≥ t0f (p∗ (tf )) Φ(tf .6). THE LINEAR OPTIMAL CONTROL PROBLEM x3 c 85 x∗ (tf ) c x2 K π ∗ = {x|c x = c x∗ (tf )} x1 Figure 7.1: c is the outward normal to π ∗ supporting K at x∗ (tf ) .5) which is equivalent to (7. ˙ ﬁnal condition: p∗ (tf ) = c .4): adjoint equation: p∗ (t) = −A (t)p∗ (t) . t0 ≤ t ≤ tf . The beauty and utility of the theory lies in the following result which translates this characterization directly in terms of u∗ .6) . for all t ∈ [t0 . t0 . Let p∗ (t) be the solution of (7. u∗ ). τ )B(τ )u(τ )dτ ] . tf ]. t (7. except possibly for a ﬁnite set. x0 . τ )B(τ )u∗ (τ )dτ ] t ≥ (p∗ (tf )) [Φ(tf . tf ∗ ∗ t0 (p (tf )) Φ(tf .1.3) (7. Theorem 1: Let u∗ (·) ∈ U and let x∗ (t) = φ(t.4) (7.3) and (7. c (x∗ (tf ) + δc) = c x∗ (tf ) + δ|c|2 > c x∗ (tf ) . t0 )x0 + t0f Φ(tf . Finally. t0 )x0 + t0f Φ(tf . ♦ The result above characterizes the optimal control u∗ in terms of the ﬁnal state x∗ (tf ). from the deﬁnition of K it follows immediately that u∗ is optimal iff c x∗ (tf ) ≥ c x for all x∈K. Then u∗ (·) is optimal iff (p∗ (t)) B(t)u∗ (t) = sup{(p∗ (t)) B(t)v|v ∈ Ω} . Proof: u∗ (·) is optimal iff for every u(·) ∈ U (p∗ (tf )) [Φ(tf . t0 ≤ t ≤ tf .

and since t∗ is a point of continuity of B(·) and u∗ (·). via the adjoint differential equation. t ∈ [t0 . and we deﬁne M by M (t. To prove the necessity let D be the ﬁnite set of points where the function B(·) or u∗ (·) is discontinuous. Indeed if this is not the case. t∗ ∈ D. We shall show that if u∗ (·) is optimal then (7.9) is the following. Taking t1 = t0 in (7.. giving a contradiction. t0 . (7. tf ∗ t0 (p (τ )) B(τ )u∗ (τ )dτ ≥ tf ∗ t0 (p (τ )) B(τ )u(τ )dτ. p∗ (t)) = M (t. Remark 2: If we deﬁne the Hamiltonian function H by H(t.7) and the sufﬁciency of (7. x. (7. t0 .7) we see that u∗ (·) cannot be optimal. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL Now by properties of the adjoint equation we know that p∗ (t)) = (p∗ (tf )) Φ(tf .86 CHAPTER 7. (7. t1 . we see that if u∗ (·) is optimal. the outward normal p∗ (tf ) at time tf . u. if c = p∗ (tf ) is the outward normal to a hyperplane supporting K(tf . x. it follows that there exists δ > 0 such that (p∗ (t)) B(t)u∗ (t) < (p∗ (t)) B(t)v. tf ] u∗ (t) otherwise .7). i. The situation is illustrated in Figure 7.8) implies that tf ∗ t0 (p (t)) (7. x0 ) at x∗ (tf ).9). and v ∈ Ω such that (p∗ (t∗ )) B(t∗ )u∗ (t∗ ) < (p∗ (t∗ )) B(t∗ )v .e. tf ].5) is satisﬁed for t ∈ D.2. This normal is obtained by transporting backwards in time. t) so that (7. x0 ) and p∗ (t) is the normal to a hyperplane supporting K(t. t0 . then there exists t∗ ∈ [t0 . p)|u ∈ Ω}. B(t)˜(t)dt > u tf ∗ t0 (p (t)) B(t)u∗ (t)dt . This condition is known as the maximum principle. x∗ (t).8) v |t − t∗ | < δ. p∗ (t)) . x∗ (t). Remark 1: The geometric meaning of (7.6) is equivalent to (7. p) = p (A(t)x + B(t)u) . p) = sup{H(t.10) . x0 ) at x∗ (t). u∗ (t). (p∗ (t2 ))x∗ (t2 ) ≥ (p∗ (t2 )) x for all x ∈ K(t2 . for |t − t∗ | < δ . x∗ (t1 )). ♦ But then from (7. Deﬁne u(·) ∈ U by ˜ u(t) = ˜ Then (7. Corollary 1: For t0 ≤ t1 ≤ t2 ≤ tf . then (7.5) is immediate.9) Exercise 2: Prove Corollary 1. then x∗ (t) is on the boundary of K(t.5) can be rewritten as H(t. x. u.

We say that p is orthogonal to T 0 at z ∗ and we write p ⊥ T 0 (z ∗ ) if .) 7. p∗ (t)) is a Lipschitz function of t. show that m(t) is constant. (Hint: Use Lemma 1 of 5. MORE GENERAL BOUNDARY CONDITIONS 87 Exercise 3: (i) Show that m(t) = M (t. t0 .e. i. { ] → ⊗ and u(·)piecewise continuous. x0 ).11) G0 and Gf are ﬁxed matrices of dimensions 0 xn and f × n respectively. + γn exp(δn (t)) for some γi .10) where p∗ (·) is the solution of the adjoint equation (7. (Hint: Show that (dm/dt) ≡ 0. while b0 ∈ R 0 . (ii) If A(t). . (7. Let f0 : Rn → R be a differentiable function and suppose that the objective function in (7.7. B(t) are constant. Show that there is an optimal control u∗ (·) and t0 ≤ t1 ≤ t2 ≤ .3) with the ﬁnal condition p∗ (tf ) = f0 (x∗ (tf )) . Maximize c x(tf ) subject to dynamics: x(t) = A(t)x(t) + B(t)u(t). (Hint: ﬁrst show that (p∗ (t)) B = γ1 exp(δ1 t) + . Also show that this condition is sufﬁcient for optimality if f0 is concave.) The next two exercises show how we can obtain important qualitative properties of an optimal control. we ﬁrst characterize optimality in terms of the state at the ﬁnal time. ≤ tn ≤ tf such that u∗ (t) ≡ α or β on [ti . t0 . β]. We will analyze the problem in the same way as before. Suppose u∗ (·) is an optimal control. Let z ∗ ∈ T 0 .2) is f0 (x(tf )) instead of c x(tf ). so that B(t) is an n × 1 matrix.) Exercise 6: Assume that K(tf .1. Exercise 5: Suppose Ω = [α.2 More General Boundary Conditions We consider the following generalization of (7. δi in R. and then translate these conditions in terms of the control. . Exercise 4: Suppose that Ω is bounded and closed. ti+1 ). Show that there exists an optimal control u∗ (·) such that u∗ (t) belongs to the boundary of Ω for all t. x∗ (t). Suppose that A(t) ≡ A and B(t) ≡ B are constant matrices and A has n real eigenvalues. x0 ) is convex (see remark in Exercise 1 above). For convenience let T 0 = {z ∈ Rn |G0 z = b0 } . ﬁnal condition: Gf x(tf ) = bf .11) In (7. Show that u∗ (·) satisﬁes the maximum principle (7. then f0x (x∗ (tf )(x∗ (tf ) − x) ≤ for all x ∈ K(tf . T f = {z ∈ Rn |Gf z = bf } . Deﬁnition: Let p ∈ Rn .2).2. ˙ initial condition: G0 x(t0 ) = b0 .1 to show that if u∗ (·) is optimal.. control constraint: u(·) ∈ U . t0 ≤ t ≤ tf . . bf ∈ R f are ﬁxed vectors. : [ . . That is. 0 ≤ i ≤ n. The notion of the previous section is retained.

t0 . x0 ) x∗ (t = K(t0 . t1 t2 tf t t0 . x0 ) K(tf . x0 ) p∗ (tf ) = c x∗ (tf ) CHAPTER 7.88 Rn Rn p∗ (t2 ) p∗ (t 1) Rn Rn x0 0) = x∗ (t1 ) K(t1 . t0 . t0 .9) for t1 = t0 . x0 ) x∗ (t2 ) K(t2 . CONTINUOUS-TIME LINEAR OPTIMAL CONTROL Figure 7.2: Illustration of (7. t0 .

t0 . (i) Suppose the Ω is convex. Secondly. then u∗ (·) is ˆ ˆ optimal and (7. not ˆ ˆ both zero. t0 )z + w|z ∈ T 0 .19) Letting r → ∞ we conclude that (7. On the other hand letting r → ˆ c x∗ (tf ) we see that (7. such that p0 r 1 + p x1 ≥ p0 r 2 + p x2 for all (r i . Also from (7. r > c x∗ (tf ). i = 1. u∗ ). z.18) implies that p0 r + p x∗ (tf ) ≥ p0 c x + p x for all x ∈ X(tf ). not both ˆ ˆ ˆ zero.18) . Similarly if z ∗ ∈ T f .16) (7. x)|r > c x∗ (tf ). If u∗ (·) is optimal. x∗ (t0 ). t0 . there exist p0 ∈ R. ˆ ˆ ˆ ˆ which is the same as (7. Exercise 1: X(tf ) = {Φ(tf . Proof: Clearly u∗ (·) is optimal iff c x∗ (tf ) ≥ c x for all x ∈ X(tf ) ∩ T f .13) are satisﬁed. u(·) ∈ U}. t0 .12) and (7. Since Ω is convex by hypothesis it follows by Exercise 1 of Section 1 that S 2 is convex. In R1+m deﬁne sets S 1 .19) can hold only if p0 ≥ 0.17) First of all S 1 ∩ S 2 = φ because otherwise there exists x ∈ X(tf ) ∩ T f such that c x > c x∗ (tf ) contradicting optimality of u∗ (·) by (7. (i) Suppose that u∗ (·) is optimal. 89 Lemma 1: Let x∗ (t0 ) ∈ T 0 and u∗ (·) ∈ U .19) can hold only if p0 c x∗ (tf ) + p x∗ (tf ) ≥ p0 c x + p x for all x ∈ X(tf ) .2. p ˆ p ˆ p ⊥ T f (x∗ (tf )) . and suppose that x∗ (tf ) ∈ T f .13) (7.15) (7. But then by the separation theorem for convex sets (see 5. x ∈ T f } . x ∈ X(tf )} .7. t0 )] (ˆ0 c + p) ⊥ T 0 (x∗ (t0 )) . x)|r = c x .12). Deﬁnition: Let X(tf ) = {Φ(tf . w ∈ K(tf . 0)}. p ∈ Rn .18) we get (7. S 2 = {(r. S 1 is convex since T f is convex.14) (ii) Conversely if there exist p0 > 0. ˆ ˆ ˆ ˆ (7. (7. S 2 by S 1 = {(r. ˆ ˆ ˆ ˆ In particular (7.12) (7. MORE GENERAL BOUNDARY CONDITIONS p (z − z ∗ ) = 0 for all z ∈ T 0 . Let x∗ (t) = φ(t. u)|z ∈ T 0 . such that (ˆ0 c + p) x∗ (tf ) ≥ (ˆ0 c + p) x for all x ∈ X(tf ) .15). p ⊥ T f (z ∗ ) if p (z − z ∗ ) = 0 for all z ∈ T f .5) there exists p0 ∈ R. p0 ≥ 0 and p ∈ Rn . and p such that (7.14) is also satisﬁed. ˆ [Φ(tf . xi ) ∈ S i . 2. p ˆ (7.20) (7.

Finally (7. we illustrate a 2-dimensional situation where T 0 = {x0 }. (7. x ∈ T f . t0 )(z − x∗ (t0 )) + x∗ (tf )} ∈ X(tf ) for all z ∈ T 0 . ˆ ˆ ˜ but then by (7. (ii) Now suppose that p0 > 0 and p are such that (7. ˆ ˆ ˜ Then from (7.15) u∗ (·) is optimal. Then in part (i) we must have p0 > 0.21) But {x − x∗ (tf )|x ∈ T f } = {z|Gf z = 0} is a subspace of Rn .12). In particular. Remark 3: In (i) the convexity of Ω is only used to guarantee that K(tf . p = (ˆ/ˆ0 ) will also satisfy (7. and T f ∩ X(tf ) consists of just one vector.12). t0 )(z − x∗ (t0 )) for all z ∈ T 0 . ˆ which is the same as (7.13).3 below. t0 . ˆ We now translate the conditions obtained in Lemma 1 in terms of the control u∗ . ♦ Remark 1: If it is possible to choose p0 > 0 then p0 = 1. 0) is convex. and (7.12) always implies (7.3) we are forced to set p0 = 0. ˆ ˆ˜ so that from (7. But it is known that K(tf . .13) we conclude that p x∗ (tf ) = p x . ˆ ˆ ˆ ˆ or p (x − x∗ (tf )) ≥ 0 for all x ∈ T f ˆ (7.13). ˆ ˆ ˆ p p (7. It follows that the control u∗ (·) ∈ U for which x∗ (tf ) = φ(tf . ˆ Remark 2: it would be natural to conjecture that in part (i) p0 may be chosen > 0. T f is the vertical line. 0) is convex even if Ω is not (see Neustadt [1963]).90 CHAPTER 7.14).13). Finally. u∗ ) ∈ T f is optimal for any c.13) are satisﬁed. In higher dimensions the reasons may be more ˆ f is “tangent” to X(t ) we may be forced to set p = 0 (see complicated. x0 . Intuitively p0 = 0 means that it is so difﬁcult to satisfy ˆ the initial and ﬁnal boundary conditions in (7.21) can hold only if p (x − x∗ (tf )) = 0 for all x ∈ T f . But in Figure ˆ 7. CONTINUOUS-TIME LINEAR OPTIMAL CONTROL p0 r + p x ≥ p0 c x∗ (tf ) + p x∗ (tf ) for all r > c x∗ (tf ). Let x ∈ X(tf ) ∩ T f . t0 . Clearly then for some c (in particular for the c in Figure 7. ˆ ˆ ˆ ˆ which can hold only if p1 c x∗ (tf ) + p x ≥ p0 c x∗ (tf ) + p x∗ (tf ) for all x ∈ T f .14) holds. Exercise 2: Suppose there exists z in the interior of X(tf ) such that z ∈ T f .14) hold for any vector c whatsoever.14). p ˆ which can hold only if (7. t0 . {Φ(tf . in part (ii) of the Lemma we may assume p0 = 1. because by the deﬁnition of X(tf ) and Exercise 1.12).12) we get 0 ≥ (ˆ0 c + p) Φ(tf .11) that optimization becomes a secondary matter. (7. so that (7. ˆ and (7. but basically if T ˆ0 f Exercise 2 below). since then (7. so that from (7. we note that part (i) is not too useful if p0 = 0.12) we get p0 c x∗ (tf ) ≥ p0 c x .

(7. u∗ ) and suppose that x∗ (tf ) ∈ T f . x. t0 . p∗ (t)) = M (t. not both identically zero. x∗ (t). x∗ (t). (7. x∗ (t0 )) ⊂ X(tf ).22) initial condition: p∗ (t0 )⊥T 0 (x∗ (t0 )) (7. and (7.14) and (7. (7. Let x∗ (t) = φ(t.24). such that (7. and a 0 function p∗ : [t0 . [Here H(t. whereas since K(tf . Next if x ∈ X(tf ) we have 0 (ˆ0 c + p) x = (p∗ (tf )) x p ˆ = (p∗ (tf )) (Φ(tf .12). t0 ≤ t ≤ tf ˙ (7.13). Let p∗ = p0 and let p∗ (·) be the solution ˆ 0 of (7. not ˆ ˆ both zero. Then by Lemma 1 there exist p ≥ 0.22). x∗ (t0 )) . 0 Then u∗ (·) is optimal. x. x.22). p) = sup{H(t. Let p0 = p∗ and p = ˆ ˆ 0 0 p∗ (tf ) − p∗ c.11). then there exist a number p∗ ≥ 0.23).22). u. v. (7.13) are respectively equivalent to (7. t0 )z + w) . (7.24) becomes equivalent to (7. If u∗ (·) is optimal for (7. 0 and the maximum principle H(t. and (7. tf ] except possibly for a ﬁnite set.7. p ∈ Rn .26) (i) Suppose u∗ (·) is optimal and Ω is convex. MORE GENERAL BOUNDARY CONDITIONS 91 Theorem 1: Let x∗ (t0 ) ∈ T 0 and u∗ (·) ∈ U .23) ﬁnal condition: (p∗ (tf ) − p∗ c)⊥T f (x∗ (tf )) .24). t0 . then (7.25) holds for all t ∈ [t0 . .25) is equivalent to (7. ˆ ˆ ˆ 0 Then (7. p)|v ∈ Ω}.23).2. u∗ (t).] Proof: A repetition of a part of the argument in the proof of Theorem 1 of Section 1 show that if p∗ satisﬁes (7. (7.26) are satisﬁed.24). p∗ (t)) .25).24) (7. tf ] → Rn . (7. so that (7. (i) Suppose that Ω is convex.22) with the ﬁnal condition p∗ (tf ) = p∗ c + p = p0 c + p .26) is implied by (7. M (t. t0 . x∗ (t0 ). satisfying adjoint equation: p∗ (t) = −A (t)p∗ (t) .26): (p∗ (tf )) x∗ (tf ) ≥ (p∗ (tf )) x for all x ∈ K(tf .14) are satisﬁed. p) = p (A(t)x + B(t)u). (ii) Suppose p∗ > 0 and (7. (7.13) and (7.23) and (7.12). (ii) Conversely suppose there exist p∗ > 0 and p∗ (·) satisfying (7.

3: Situation where p0 = 0 ˆ x0 = T 0 c t . CONTINUOUS-TIME LINEAR OPTIMAL CONTROL Tf . x0 ) = X(tf ) f f Tf Figure 7. t0 . x (t ) = X(t ) ∗ K(tf .92 CHAPTER 7.

t0 . t0 )x∗ (t0 )) ∈ K(tf . t0 )x∗ (t0 )) . x∗ (t0 )). it follows from (7. Thus (ˆ0 c + p) x∗ (tf ) ≥ (ˆ0 c + p) x for all x ∈ X(tf ) . t0 )(z − x∗ (t0 )) p ˆ +(p∗ (tf )) (w + φ(tf . 0). Hence (ˆ0 c + p) x = (p∗ (f )) Φ(tf . Exercise 4: How would you use Exercise 3 to solve Example 3 of Chapter 1? . M (t. t0 . p ˆ p ˆ and so u∗ (·) is optimal by Lemma 1. in (7. t0 )x∗ (t0 )) = (p∗ (t0 )) (z − x∗ (t0 )) +(p∗ (tf )) (w + Φ(tf . and since (w+φ(tf .7. ♦ Exercise 3: Suppose that the control constraint set is Ω(t) which varies continuously with t. 93 But by (7. p)|v ∈ Ω(t)}.25). and we require that u(t) ∈ Ω(t) for all t. Show that Theorem 1 also holds for this case where. x. x. p) =sup{H(t. MORE GENERAL BOUNDARY CONDITIONS for some z ∈ T 0 and some w ∈ K(tf .23) the ﬁrst term on the right vanishes.2.26) that the second term is bounded by (p∗ (tf )) x∗ (tf ). v.

CONTINUOUS-TIME LINEAR OPTIMAL CONTROL .94 CHAPTER 7.

et al. But ﬁrst we need to impose some regularity conditions on the differential equation (8. 8. Finally. We assume throughout that the function f : [t0 . (For complete proofs see (Lee and Markus [1967] or Pontryagin. Instead we shall settle for a comparison between the trajectory x∗ (·) and trajectories x(·) obtained by perturbing the control u∗ (·) and the initial condition x∗ (t0 ). However.) The principal result. ˙ Rn Rp u∗ (·) (8. We can then estimate the difference between x(·) and x∗ (·) by the solution to a linear differential equation as shown in Lemma 1 below.1 Main Results 8.. u(t)) .1. We are interested in the optimal control of a system whose dynamics are governed by the nonlinear differential equation x(t) = f (t. it is possible to convey the main ideas of the proofs at an intuitive level and we shall do so. Unfortunately when f is nonlinear such a characterization is not available. In the case of linear systems we obtained the necessary and x conditions for optimality by comparing x∗ (·) with trajectories x(·) corresponding to other admissible controls u(·).2 is presented in Section 1.Chapter 8 SEQUENTIAL DECISION PROBLEMS: CONTINUOUS-TIME OPTIMAL CONTROL OF NONLINEAR SYSTEMS We now present a sweeping generalization of the problem studied in the last chapter. in Section 5 we discuss the so-called singular case and also analyze Example 4 of Chapter 1. This comparison was possible because we had an explicitly characterization of x(·) in terms of u(·). Suppose is an optimal control ∗ (·) is the corresponding trajectory. An alternative form of the objective function is discussed in Section 2. Unfortunately we are forced to omit the proofs of the results since they require a level of mathematical sophistication beyond the scope of these Notes. Section 3 deals with the minimum-time problem and Section 4 considers the important special case of linear systems with quadratic cost. t0 ≤ t ≤ tf .1).1 Preliminary results based on differential equation theory. tf ] × Rn × Rp → Rn satisﬁes the following conditions: 95 . which is a direct generalization of Theorem 1 of 7. (t). x. [1962].1) where x(t) ∈ is the state and u(t) ∈ is the control.

i D. t1 . . tm < tf . Theorem 1: For every z ∈ Rn . m. fx . i = 1. (t). and every piecewise continuous function u(·) : [t0 . . ·) : Rn xRp → Rn is continuously differentiable in the remaining variables (x. . m (recall that D is the set of ≥ 0. u). z. . for every ﬁnite α. there exist ﬁnite number β and γ such that |f (t. z. . . for each ﬁxed t ∈ [t0 . x ∈ Rn . i = 1. Let u∗ (·) ∈ U be ﬁxed and let D∗ be the set of discontinuity points of u∗ (·). . the n × n matrix-valued function Φ deﬁned by Φ(t2 . 0 Deﬁnition: π = (t1 . t1 ≤ t ≤ tf . t1 ≤ t ≤ tf . u(t)) . u(·)) : Rn → Rn is differentiable. ∂x and the initial condition Φ(t1 . z. tf ] → Ω. . t1 . m. except for a ﬁnite subset D ⊂ [t0 . CONINUOUS-TIME OPTIMAL CONTROL 1. there exists a unique solution x(t) = φ(t. u(·)) . t1 ≤ t ≤ tf . for ﬁxed t1 ≤ t2 in [t0 . the function φ(t2 . u ∈ Rp with |u| ≤ α . the functions f. The following result is proved in every standard treatise on differential equations. u(·)) = In . 2. ˙ satisfying the initial condition x(t1 ) = z . tf ]. tf ]. tf ]. t1 . and 4. z. . and 3. . x(t). um ) is said to be a perturbation data for u∗ (·) if 1. 1. t1 . u(·)) is the solution of the linear homogeneous differential equation ∂Φ ∂t (t. x. m. and ti ∈ D ∗ discontinuity points of f ). Moreover. t0 < t1 < t2 < . . u(t))]Φ(t. . ·. . . . . u)| ≤ β + γ|x| for all t ∈ [t0 . z. t1 . tf ]. z. of the differential equation x(t) = f (t. . . tf ] and ﬁxed u(·). u(·)) = ∂φ ∂z (t2 . 2. x. i = 1. . t1 . . ui ∈ Ω. . for every t1 ∈ [t0 . fu are continuous on [t0 . tf ]×Rn × Rp . Furthermore. . . . 3. f (t. ·. (·)) = [ ∂f (t. tf ] → Rp . . m is a nonnegative integer. u1 . tm . . t1 . Let x∗ ∈ Rn be a ﬁxed initial condition. u(·)). . u.96 CHAPTER 8. Now let Ω ⊂ Rp be a ﬁxed set and let U be set of all piecewise continuous functions u(·) : [t0 .

and it is omitted (see for example (Lee and Markus [1967])).ε) (·) ∈ U corresponding to π is deﬁned by u(π. uj ) − f (tj . Then for 0 ≤ ε ≤ ε(π). tj ] = φ for i = j.0) (·)is the linearized perturbation corresponding to(π. tj )[f (tj . tf ] . t1 ) 1 . Deﬁnition: For each t ∈ [t0 . and [ti −ε i . u∗ (·)). x(ξ. u(π. x(ξ. ti ] . t0 .ε) ) up to an error of order o(ε). x∗ . t0 . Remark: By Lemma 1 (x∗ (t)+εh(π. = Φ(t. x∗ ). u∗ (t1 ))] i . ti ] ⊂ [t0 . x∗ (tj ). m u∗ (t) otherwise . t0 )ξ = Φ(t. ξ). t1 ) = 0 Φ(t2 . tf ] for all i.1. . z) = {φ(t. x∗ (tj ). .ε) (t) = Φ(t. 0)} . t0 )ξ + j=1 j (See Figure 8. x∗ (tj ). . t0 . u(·))|u(·) ∈ U} be the set of states reachable at time t.ε) = x∗ . Deﬁnition: Any vector ξ ∈ Rn is said to be a perturbation for x∗ . t1 . .ε) . t ∈ [t0 . x∗ (t1 ). x∗ (t1 ). starting at time t0 in state z.ε) − x∗ ) = ξ . t0 )ξ + Φ(t. Lemma 1: lim |xε (t) − x∗ (t) − εh(π.ξ) (·) the linearized (trajectory) perturbation corresponding to (π. u∗ (tj ))] Φ(t. where h(π. In particular for ξ = 0. let Q(t) = {h(π. and using controls u(·) ∈ U . MAIN RESULTS 97 Let ε(π) > 0 be such that for 0 ≤ ε ≤ ε(π) we have [ti − ε i . i = 1.) We call h(π. t ∈ [tm . u1 ) − f (t1 . t ∈ [t1 . Let Φ(t2 . tf ]. x∗ (t1 ).ε) (t)| = 0 for t ∈ [t0 . u∗ (·)) and let xε (t) = φ(t. 0 More precisely we have the following result which we leave as an exercise. tf ] let K(t. t ∈ [t0 . 0 ε→0 and ε→0 lim 1 ε (x(ξ. z. ti+1 ) . t0 )ξ + j=1 m Φ(t.ε) (t) = ui for all t ∈ [ti − ε i . ti ] [tj −ε j .the perturbed control u(π. t ∈ [ti . t0 .0) (t)|πis a perturbation data for u∗ (·). the set x∗ (t) + Q(t) can serve as an approximation to the set K(t. . t2 ) j = Φ(t. Deﬁnition: For z ∈ Rn .ξ) ) belongs to the set K(t. t0 . t0 .8. The following lemma gives an estimate of x∗ (t) − xε (t). 0 Now let x∗ (t) = φ(t. t1 )[f (t1 .ε) (·)). tm )[f (tj . The proof of the lemma is a straightforward exercise in estimating differences of solutions to differential equations. x∗ (tj ).ε) (·) is given by ε→0 h(π. and a function x(ξ. t1 ]. and h(π.ε) deﬁned for 0 ε > 0 is said to be a perturbed initial condition if lim x(ξ. u∗ tj ))] .1. uj ) − f (tj .

control constraint: u(·) ∈ U .1.) Show that Q(t) ⊂ C(K(t. be the solution of (8. t0 ≤ t ≤ tf . ˙ initial condition: x(t0 ) = x∗ . CONINUOUS-TIME OPTIMAL CONTROL u1 u(πε) (·) u∗ (·) ε | 1 u2 t1 | | u3 ε 2 ε | | 3 t0 x t2 t3 tf x∗ (·) | | xε (·) x( ξ. be the 0 corresponding trajectory. ˙ ∂x (8. x∗ ). Let p∗ (t).5): adjoint equation: p∗ (t) = −[ ∂f (t. x∗ (t). t0 ≤ t ≤ tf . u : [t0 .4) and (8. t0 ≤ t ≤ tf .1: Illustration for Lemma 1. t0 . u(t)) . (8. x∗ .1.3) where ψ : Rn → R is differentiable and f satisﬁes the conditions listed earlier.1.2) (8. x(t). t0 .e.3): Maximize ψ(x(tf )) subject to dynamics: x(t) = f (t.4) . t0 ≤ t ≤ tf . Let u∗ (·) ∈ U be an optimal control and let x∗ (t) = φ(t.98 u CHAPTER 8. tf ] → Ω and u(·) piecewise continuous . i. Theorem 2: Consider the optimal control problem (8. u∗ (t))] p∗ (t). 0 ﬁnal condition: x(tf ) ∈ Rn . x∗ (t)) . u∗ (·)).. ε) εhπξ t1 | | | | t2 t3 tf Figure 8. Exercise 1: (Recall the deﬁnition of the tangent cone in 5. 0 We can now prove a generalization of Theorem 1 of 7.

0) (t∗ ) > 0 . t0 . Then for all ε > 0 sufﬁciently small. asserts further that if h is in the interior of C(t) then in fact (x∗ (t) + εh) ∈ K(t.1 ψ(x∗ (tf ))h ≤ 0 for all h ∈ C(K(tf . 1. But ﬁrst we need some simple properties of the sets Q(t) which we leave as exercises. p∗ (t∗ ) [f (t∗ .7) Now suppose that (8.1. v). Lemma 2: Let h belong to the interior of the cone C(t).1. In Theorem 2 the initial condition is ﬁxed and the ﬁnal condition is free. Also h(π.4) we can see that p∗ (t∗ ) = p∗ (tf ) Φ(tf . (8. Exercise 2: Show that (i) Q(t) is a cone. Instead we offer a plausibility argument. x. Deﬁnition: Let C(t) denote the closure of Q(t).0) (tf ) = Φ(tf . u.2) ψx (x∗ (tf ))h ≤ 0 for all h ∈ Q(tf ) . Then there exists v ∈ Ω such that (8.2 More general boundary conditions. Exercise 3: Show that (i) C(t) is a convex cone. t0 . then (8. M (t. Proof: Since u∗ (·) is optimal we must have ψ(x∗ (tf )) ≥ ψ(z) for all z ∈ K(tf . p) = p f (t. x.e. x(t∗ ). u∗ (t). 99 (8..9) Now from (8.8) is equivalent to p∗ (t∗ ) h(π. MAIN RESULTS ﬁnal condition: p∗ (tf ) = Then u∗ (·) satisﬁes the maximum principle H(t. t0 . [Here H(t. Lemma 1 needs to be extended to Lemma 2 below.6) for all t ∈ [t0 .6) does not hold from some t∗ ∈ D ∗ ∪ D. x∗ (t0 )) up to an error of order o(ε). v. x∗ ) . v) − f (t∗ .7). x. p) = sup{H(t. . i. below. x∗ (t). (ii) for t0 ≤ t1 ≤ t2 ≤ tf . p)|v ∈ Ω}]. The proof of the lemma depends upon some deep topological results and is omitted. 0 and in particular from (8. Lemma 2. x∗ (t0 )) for ε > 0 sufﬁciently small. Φ(t2 . x. tf ] except possibly for a ﬁnite set. ♦ (8.9) is equivalent to p∗ (tf ) h(π. if h ∈ Q(t) and λ ≥ 0. t∗ )h(π. x∗ ).8. t1 )Q(t1 ) ⊂ Q(t2 ) . p∗ (t)) ψ(x∗ (tf )) .0) (tf ) > 0 which contradicts (8. t0 . If we consider the perturbation data π = (t∗ . The problem involving more general boundary conditions is much more complicated and requires more reﬁned analysis. x∗ (t).8) 8. Speciﬁcally. x(t∗ ). u∗ (t∗ ))] > 0 . ). Remark: From Lemma 1 we know that if h ∈ C(t) then (x∗ (t) + εh) belongs to K(t. u. Φ(t2 . (ii) for t0 ≤ t1 ≤ t2 ≤ tf . x∗ (tf )) .1. 0 and so by Lemma 1 of 5. t1 )C(t1 ) ⊂ C(t2 ) .5) (8. p∗ (t)) = M (t.0) (t∗ ) so that (8. t∗ ). then λh ∈ Q(t).

100

CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
(x∗ (t) + εh) ∈ K(t, t0 , x∗ ) . 0 (8.10)

Plausibility argument. (8.10) is equivalent to εh ∈ K(t, t0 , x∗ (t0 )) − {x∗ (t)} , where we have moved the origin to x∗ (t). The situation is depicted in Figure 8.2. ˆ K(ε) ˆ C(ε) o(ε) K(t1 , t0 , x∗ ) − {x∗ (t)} εh 0

(8.11)

h

δε

C(t)

Figure 8.2: Illustration for Lemma 2. ˆ Let C(ε) be the cross-section of C(t) by a plane orthogonal to h and passing through εh. Let ˆ K(ε) be the cross-section of K(t, t0 , x∗ ) − {x∗ (t0 )} by the same plane. We note the following: 0 ˆ ˆ (i) by Lemma 1 the distance between C(ε) and K(ε) is of the order o(ε); ˆ (ii) since h is in the interior of C(t), the minimum distance between εh and C(ε) is δε where δ > 0 is independent of ε. ˆ Hence for ε > 0 sufﬁciently small εh must be trapped inside the set K(ε). (This would constitute a proof except that for the argument to work we need to show that there ˆ are no “holes” in K(ε) through which εh can “escape.” The complications in a rigorous proof arise precisely from this drawback in our plausibility argument.) ♦ ∗ ) in a neighborhood of x∗ (t) when we Lemmas 1 and 2 give us a characterization of K(t, t0 , x0 perturb the control u∗ (·) leaving the initial condition ﬁxed. Lemma 3 extends Lemma 2 to the case when we also allow the initial condition to vary over a ﬁxed surface in a neighborhood of x∗ . 0 0 Let g0 : Rn → R 0 be a differentiable function such that the 0 × n matrix gx (x) has rank 0 n 0 0 0 ∗ 0 0 for all x. Let b ∈ R be ﬁxed and let T = {x|g (x) − b }. Suppose that x0 ∈ T and let 0 (x∗ ) = {ξ|g 0 (x∗ )ξ = 0}. Thus, T 0 (x∗ ) + {x∗ } is the plane through x∗ tangent to the surface T x 0 0 0 0 0 T 0 . The proof of Lemma 3 is similar to that of Lemma 2 and is omitted also. Lemma 3: Let h belong to the interior of the cone {C(t)+Φ(t, t0 )T 0 (x∗ )}. For ε ≥ 0 let h(ε) ∈ Rn 0 1 be such that lim h(ε) = 0, and lim ( )h(ε) = h. Then for ε > 0 sufﬁciently small there exists ε→0 ε x0 (ε) ∈ T 0 such that (x∗ (t) + h(ε)) ∈ K(t, t0 , x0 (ε)) .

8.1. MAIN RESULTS

101

We can now prove the main result of this chapter. We keep all the notation introduced above. f Further, let gf : Rn → R f be a differentiable function such that gx (x) has rank f for all x. f ∈ Rn be ﬁxed and let T f = {x|g f (x) − bf }. Finally, if x∗ (t ) ∈ T f let T f (x∗ (t )) = Let b f f f {ξ|gx (x∗ (tf ))ξ = 0}. Theorem 3: Consider the optimal control problem (8.12): Maximize ψ(x(tf )) subject to dynamics: x(t) = f (t, x(t), u(t)) , t0 ≤ t ≤ tf , ˙ initial conditions: g0 (x(t0 )) = b0 , ﬁnal conditions: gf (x(tf )) = bf , control constraint: u(·) ∈ U , i.e., u : [t0 , tf ] → Ω and u(·) piecewise continuous .

(8.12)

Let u∗ (·) ∈ U , let x∗ ∈ T 0 and let x∗ (t) = φ(t, t0 , x∗ , u∗ (·)) be the corresponding trajectory. 0 0 Suppose that x∗ (tf ) ∈ T f , and suppose that (u∗ (·), x∗ ) is optimal. Then there exist a number 0 p∗ ≥ 0, and a function p∗ : [t0 , tf ] → Rn , not both identically zero, satisfying 0 adjoint equation: p∗ (t) = −[ ∂f (t, x∗ (t), u∗ (t))] p∗ (t), t0 ≤ t ≤ tf , ˙ ∂x initial condition: p∗ (t0 )⊥T 0 (x∗ ) , 0 ﬁnal condition: (p∗ (tf ) − p∗ ψ(x∗ (tf )))⊥T f (x∗ (tf )) . 0 Furthermore, the maximum principle H(t, x∗ (t), u∗ (t), p∗ (t)) = M (t, x∗ (t), p∗ (t)) (8.16) (8.13) (8.14) (8.15)

holds for all t ∈ [t0 , tf ] except possibly for a ﬁnite set. [Here H(t, x, p, u) = p f (t, x, u, ), M (t, x, p) = sup{H(t, x, v, p)|v ∈ Ω}]. Proof: We break the proof up into a series of steps. Step 1. By repeating the argument presented in the proof of Theorem 2 we can see that (8.15) is equivalent to p∗ (tf ) h ≤ 0 for all h ∈ C(tf ) . (8.17)

Step 2. Deﬁne two convex sets S1 , S2 in R1+m as follows: S1 = {(y, h)|y > 0, h ∈ T f (x∗ (tf ))}, S2 = {(y, h)|y = ψx (x∗ (tf ))h, h ∈ {C(tf ) + Φ(tf , t0 )T 0 (x∗ )}} . 0 We claim that the optimality of (u∗ (·), x∗ ) implies that S1 ∩ Relative Interior (S2 ) = φ. Suppose 0 this is not the case. Then there exists h ∈ T f (x∗ (tf )) such that ψx (x∗ (tf ))h > 0 , (8.18)

102

CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL
h ∈ Interior{C(tf ) + Φ(tf , t0 )T 0 (x∗ )} . 0 (8.19)

f f Now by assumption gx (x∗ (tf ) has maximum rank. Since gx (x∗ (tf ))h = 0 it follows that the Implicit Function Theorem that for ε > 0 sufﬁciently small there exists h(ε) ∈ Rn such that

gf (x∗ (tf ) + h(ε)) = bf ,

(8.20)

and, moreover, h(ε) → 0, (1/ε)h(ε) → h as ε → 0. From (8.18) and Lemma 3 it follows that for ε > 0 sufﬁciently small there exists x0 (ε) ∈ T 0 and uε (·) ∈ U such that x∗ (tf ) + h(ε) = φ(tf , t0 , x0 (ε), uε (·)) . Hence we can conclude from (8.20) that the pair (x0 (ε), uε (·)) satisﬁes the initial and ﬁnal conditions, and the corresponding value of the objective function is ψ(x∗ (tf ) + h(ε)) = ψ(x∗ (tf )) + ψx (x∗ (tf ))h(ε) + o(|h(ε)|) , and since h(ε) = εh + o(ε) we get ψ(x∗ (tf ) + h(ε)) = ψ(x∗ (tf )) + ε)ψx (x∗ (tf ))h + o(ε) ; but then from (8.18) ψ(x∗ (tf ) + h(ε)) > ψ(x∗ (tf )) for ε > 0 sufﬁciently small, thereby contradicting the optimality of (u∗ (·), x∗ ). 0 Step 3. By the separation theorem for convex sets there exist p0 ∈ R, p1 ∈ Rn , not both zero, such ˆ ˆ that p0 y 1 + p1 h1 ≥ p0 y 2 + p1 h2 for all (y i , hi ) ∈ S1 , i = 1, 2 . ˆ ˆ ˆ ˆ (8.21)

Arguing in exactly the same fashion as in the proof of Lemma 1 of 7.2 we can conclude that (8.21) is equivalent to the following conditions: p0 ≥ 0 , ˆ p1 ⊥T f (x∗ (tf )) , ˆ Φ(tf , t0 ) (ˆ0 ψ(x∗ (tf )) + p1 )⊥T 0 (x∗ ) , p ˆ 0 and (ˆ0 ψx (x∗ (tf )) + p1 )h ≤ 0 for all h ∈ C(tf ) . p ˆ (8.24)

(8.22)

(8.23)

If we let p∗ = p0 and p∗ (tf ) = p0 ψ(x∗ (tf )) + p1 then (8.22), (8.23), and (8.24) translate respecˆ0 ˆ ˆ ˆ tively into (8.15), (8.14), and (8.17). ♦

x. gf (x) = bf are augmented g0 (˜) = ˜ x x0 g0 (x) = ˜0 = b 0 b0 and gf (˜) = gf (x) = bf . p∗ (t)) ≡ constant. x) ∈ ˜ R1+m as follows: · f0 (t. x. then there exists a function p∗ = (p∗ . Evidently then the problem of maximizing (8. 0 ﬁnal condition: p∗ (tf )⊥T f (x∗ (tf )) . x∗ (t). u(t))dt (8. v)|v ∈ Ω}. let x∗ ∈ T o and let x∗ (t) = φ(t. x= ˜ = f (t.2.2 Integral Objective Function In many control problems the objective function is not given as a function ψ(x(tf )) of the ﬁnal state. x∗ (t). u(t))dt . x∗ ) is optimal. u∗ (t)) = M (t.) ˜ · ˜ . the boundary conditions. We proceed to show how such objective functions can be treated as a special case of the problems of the last section. ˙ initial conditions: g0 (x(t0 )) = b0 . x(t). p. then M ˜ ˜ Exercise 1: Prove Theorem 1. but rather as an integral of the form tf t0 f0 (t. u(t)) = x(t) ˙ f (t. p∗ ) : [t0 . (Hint: For the ﬁnal part show that (d/dt) M (t. satisfying 0 0 (augmented) adjoint equation: p∗ (t) = −[ ∂ f (t. t0 ≤ t ≤ tf . tf ] except possibly for a ﬁnite set. x∗ . the maximum principle ˜ ˜ H(t.25) is equivalent to the ˜ x problem of maximizing ψ(˜(tf )) = x0 (tf ) . x(t). Finally. Let u∗ (·) ∈ U. control constraint: u(·) ∈ U . p∗ (t)) ˜ ˜ ˜ holds for all t ∈ [t0 . u). u(t)) x0 (t) ˙ ˜ ˜ . and control constraints are the same as before. u∗ (·)). if f0 and f do not explicitly depend on t. x∗ (t). not identically ˜ 0 0 zero. t0 . x subject to the augmented dynamics and constraints which is of the form treated in Theorem 3 of Section 1. u(t)) The initial and ﬁnal conditions which are of the form g0 (x) = b0 . u) = p f (t. x. x. and suppose that x∗ (tf ) ∈ T f . x(t). and with p∗ (t) ≡ constant and p∗ (t) ≥ 0.] ˜ ˜ ˜ (t. u) + p f (t. x. and M (t. x. x(t). To this end we deﬁned the augmented system with state variable x = (x0 . u∗ (t))] p∗ (t) . x(t).26): tf Maximize t0 f0 (t. Theorem 1: Consider the optimal control problem (8.8. Futhermore. p. and we get the following result. u) = ˜ ˜˜ ˜ ˜ p0 f0 (t. INTEGRAL OBJECTIVE FUNCTION 103 8. [Here H(t. ˜ ˜ ∂x ˜ initial condition: p∗ (t0 )⊥T 0 (x∗ ) . x(t). tf ] → R1+m . x∗ (t). p∗ (t). (8. p) = sup{H(t. ﬁnal conditions: gf (x(tf )) = bf .26) subject to dynamics: x(t) = f (t. If 0 0 (u∗ (·). x∗ (t). u(t)). p∗ (t)) ≡ 0.25) The dynamics of the state.

28) by x(t) = z(s(t)) .27) by converting the variable time interval [t0 . ﬁnal condition: gf (x(t)f )) = bf . tf ] into a ﬁxed-time interval [0. z(s). 1]. t0 ≤ t ≤ tf .1 Main result. In many important cases the ﬁnal time is itself a decision variable. then it is easy to see that z(·) is the solution of dz ds (s) (8. In the problem considered up to now the ﬁnal time tf is assumed to be ﬁxed. v(s)) . s(·) is the solution of the differential equation s(t) = 1/α(s(t)). . CONINUOUS-TIME OPTIMAL CONTROL 8.27).29) we can obtain the solution x(·) of (8. u(t)) . 0 ≤ s ≤ 1 . t0 ≤ t ≤ tf . s(t0 ) = 0. 0 ≤ s ≤ 1 z(0) = x0 . 1] is the functional inverse of s(t). x. control constraint: u(·) ∈ U . x(t0 ) = x0 ˙ and if we deﬁne z(s) = x(t(s)). (8. Here α(s) is a new control variable constrained by α(s) ∈ (0. (8. ﬁnal-time constraint: tf ∈ (t0 . 1]. tf Maximize t0 f0 (t. v(s) = u(t(s)) . x(t). with initial condition t(0) = t0 .27) We analyze (8. 0 ≤ s ≤ 1 . consider the optimal control problem (8. in fact.3 Variable Final Time 8.28) = α(s)f (s. tf ] → [0. x(t). ˙ initial condition: g0 (x(t0 )) = b0 . where s(·) : [t0 .29) Conversely from the solution z(·) of (8. u(t)). (t). ∞).104 CHAPTER 8. ˙ . This change of time-scale is achieved by regarding t as a new state variable and selecting a new time variable s which ranges over [0. t0 ≤ t ≤ tf .3. u(t))dt subject to dynamics: x(t) = f (t. The equation for t is dt(s) ds = α(s) . ∞) . More generally. One such case is the minimum-time problem where we want to transfer the state of the system from a given initial state to a speciﬁed ﬁnal state in minimum time. Now if x(·) is the solution of x(t) = f (t.

x∗ . 0 ·∗ ˜ (8. t∗ ∈ (t0 . and ˜ 0 f with p∗ (t) ≡ constant and p∗ (t) ≥ 0. t(s)) = (f (t(s). t(1) ∈ R . z ∗ (·)) satisﬁes the ﬁnal conditions of (8. t0 ≤ t ≤ t∗ . z(s). t∗ ] → R1+m . f Exercise 1: Prove Lemma 1. t0 . t0 ≤ t ≤ tf .30). α∗ (·)). ∞) for 0 ≤ s ≤ 1 and v(·). Deﬁne x∗ . Theorem 1: Let u∗ (·) ∈ U. α∗ (·)) be an admissible control for (8.8. not identically zero. z0 ) is optimal for (8. and suppose that x∗ (t∗ ) ∈ T f . v(s))α(s)ds (8. z) ∈ R1+m . p∗ ) : [t0 .3. t∗ ) is optimal for (8.27). ∗ (ii) Let z0 ∈ T 0 . f t∗ = t∗ (1) . f ∗ Then ((v ∗ (·). u∗ (·) ∈ U. u∗ (t))] p∗ (t) . x∗ (t). z(s). α∗ (·)). ˜ ˜ ∂x ˜ initial condition: p∗ (t0 )⊥T 0 (x∗ ) . u∗ (·)). and suppose that (u∗ (·). 0≤s≤1. z0 ) is optimal for (8. 0 u∗ (t) = v ∗ (s∗ (t)) . control constraint: (v(s). If (u∗ (·).32) . and α∗ (·) by ∗ z0 = x∗ 0 ∗ (s) = u∗ (t + s(t∗ − t )) v . 0 f0 (t(s). and let 0 f x∗ (t) = φ(t. Deﬁne z0 . where the state vector (t. u∗ (·)) be the 0 0 f corresponding trajectory. t∗ ) is optimal 0 0 f f for (8.30) Lemma 1: (i) Let x∗ ∈ T 0 . then there exists a function p∗ = (p∗ . z0 . t0 . ∞).30) is established in the following result.30). x∗ . satisfying 0 0 (augmented) adjoint equation: p (t) = −[ ∂ f (t. let x∗ ∈ T 0 .27). and let (v ∗ (·). The relation between problems (8. ﬁnal constraint: gf (z(1)) = bf . α(s)). v ∗ (·).30). x∗ . 0≤s≤1. v(s))α(s). v) ∈ R1+p : 1 Maximize subject to ˙ dynamics: (z(s).30) such that the correspond∗ ing trajectory (t∗ (·).31) (8. u∗ (·) ∈ U .30). and the control (α. Suppose that ((v ∗ (·). x∗ . t∗ ) is optimal for 0 f f ∗ (8. 0 0 f α∗ (s) = (t∗ − t0 ) . α(·) piecewise continuous. f ∗ where s∗ (·) is functional inverse of t∗ (·).27). VARIABLE FINAL TIME 105 With these ideas in mind it is natural to consider the ﬁxed-ﬁnal-time optimal control problem (8. ∞) and let x∗ (t) = φ(t. Then (u∗ (·). let t∗ ∈ (0. α(s)) ∈ Ω × (0.27) and (8. t(0) = t0 . Suppose that x∗ (t∗ ) ∈ T f . ˙ initial constraint: g0 (z(0)) = b0 . and t∗ by 0 f ∗ x∗ = z0 .

so that in particular f f z ∗ (1) = x∗ (t∗ ). x∗ (t∗ ). p∗ (t)) ≡ 0. p∗ (t)) . and with λ0 0 (8. λ∗ ≡ constant. λ∗ (1) = 0 . n+1 Furthermore.40) we have (λ∗ . z0 = x∗ .38) ﬁnal condition: λ∗ (1)⊥T f (z ∗ (1)) . p∗ (t∗ ). there exists a function λ∗ = (λ∗ . f f (8. λ∗ (1) = 0 so that we would have λ∗ ≡ 0 n+1 n+1 . and deﬁne p∗ = (p∗ . not 0 n+1 ∗ (s) ≡ constant and λ∗ (s) ≥ 0. f f ˜ f f ˆ Finally. v ∗ (s))α∗ (s) 0 +λ∗ (s) f (t∗ (s).40) First of all. z ∗ (s). Because if p∗ ≡ 0. v ∗ (s))] λ∗ (s)}α∗ (s) adjoint equation:    ∂z   ∂f0 ∗  λ∗ (t)  ˙  {[ ∂t (t (s). x∗ (t). but from (8. t∗ must be such that f ˆ H(t∗ . v ∗ (s))] λ∗ (s) n+1 0 +[ ∂f (t∗ (s). f ˜ By Theorem 1 of Section 2. u∗ (t)) = M (t. z ∗ (s). tf ] except possibly for a ﬁnite set. z ∗ (s). p∗ (t) = λ∗ (s∗ (t)). t0 ≤ t ≤ t∗ .33) Also the maximum principle ˜ ˜ H(t. ˜ ∗ Proof: By Lemma 1.39) holds for all s ∈ [0. λ∗ . then from (8. v ∗ (s))] λ∗ (s) ˙  λ∗ (t)   0 ∂z    = −  +[ ∂f (t∗ (s).37) (8. v ∗ (s) = u∗ (t0 + s(t∗ − t0 )) and α∗ (s) = (t∗ − t0 ) for 0 ≤ s ≤ 1 0 f f constitute an optimal solution for (8. t∗ (s) = t0 + s(t∗ − t0 ). The resulting trajectory is z ∗ (s) = x∗ (t0 + s(t∗ − t0 )). z ∗ (s). z ∗ (s). p∗ is not identically zero. Furthermore.30). CONINUOUS-TIME OPTIMAL CONTROL ﬁnal condition: p∗ (t∗ )⊥T f (x∗ (t∗ )) . 1] except possibly for a ﬁnite set. p∗ (t). w)β 0 +λ∗ (s) f (t∗ (s). ∞)} n+1 (8. w)β + λ∗ (s)β]|w ∈ Ω.106 CHAPTER 8. Let s∗ (t) = (t − t0 )/(t∗ − t0 ). z ∗ (s). 0 ≤ s ≤ 1 .34) (8. if f0 and f do not explicitly depend on t. 1] → R1+n+1 . x∗ (t). 0 0 f (8. λ∗ ) : [0. satisfying identically zero. v ∗ (s))] λ∗ (s)}α∗ (s) ∂t  ∗ initial condition: λ∗ (0)⊥T 0 (z0 )        (8.36).35)   ˙ 0 λ∗ (t) 0  {[ ∂f0 (t∗ (s).38). x∗ (t). λ∗ ) ≡ 0 and ˜ ˜ 0 ˜ then from (8.36) (8. t∗ ] → R1+n by ˜ 0 f f f p∗ (t) = λ∗ (s∗ (t)). z ∗ (s). z ∗ (s). the maximum principle λ∗ (s)f0 (t∗ (s). t0 ≤ t ≤ t∗ . β ∈ (0. then M (t. v ∗ (s))α∗ (s) + λ∗ (s)α∗ (s) n+1 = sup{[λ∗ (s)f0 (t∗ (s). p∗ ) : [t0 . u∗ (t∗ )) = 0 . ˜ ˜ holds for all t ∈ [t0 .

Then there exists a function p∗ : [t0 . x∗ (t). z ∗ (s).3. p∗ (tf )) ≥ 0 f and if f does not depend explicitly on t then M (t. ♦ 8.35) and the fact that M (t. Theorem 2: Let t∗ ∈ (t0 .37) ˜ and (8. x(t).8. the last assertion of the Theorem follows from (8. p∗ (t)) ≡ ˜ constant if f0 . p∗ (t)) holds for all t ∈ [t0 . w) + λ∗ (s) f (t∗ (s). u∗ (t)) = M (t.3. v ∗ (s)) + λ∗ (s) = 0 n+1 and λ∗ (s)f0 (t∗ (s). v ∗ (s)) 0 = Sup {[λ∗ (s)f0 (t∗ (s).27): tf Maximize t0 (−1)dt subject to dynamics: x(t) = f (t. (8. ﬁnal condition: x(tf ) = xf . Applying Theorem 1 to this problem gives Theorem 2. f are not explicitly dependent on t.41) Evidently (8. x∗ (t).42) (8. Let x∗ (·) be the corresponding f f trajectory.39) is equivalent to λ∗ (s)f0 (t∗ (s). z ∗ (s). VARIABLE FINAL TIME 107 which is a contradiction.31). f Also the maximum principle H(t. f Finally. t∗ ] → Rn .35) follows from (8. x∗ (t). z ∗ (s). so that the optimal control problem consists of ﬁnding a control which transfers the system from state x0 at time t0 to state xf in minimum time.42) is equivalent to (8. ﬁnal-time constraint: tf ∈ (t0 .38) respectively imply (8. 0 (8. x0 .43) In (8.43). p∗ . x∗ (tf ).41) and the fact that λ∗ (1) = 0. w)]|w ∈ Ω}. It is trivial to verify that p∗ (·) satisﬁes (8. and. ∞) . u∗ (t))] p∗ (t). t0 ≤ t ≤ t∗ . Next.33). M (t∗ .45) (8.34) and (8. xf are ﬁxed. ˙ f ∂x initial condition: p∗ (t0 ) ∈ Rn . ﬁnal condition: p∗ (t∗ ) ∈ Rn . We consider the following special case of (8. x∗ (t). n+1 ˜ Finally. z ∗ (s).44) . satisfying f adjoint equation: p∗ (t) = −[ ∂f (t.46) (8. z ∗ (s). v ∗ (s)) + λ∗ (s) f (t∗ (s). t0 ≤ t ≤ tf ˙ initial condition: x(t0 ) = x0 . t∗ ] except possibly for a ﬁnite set. t∗ ] → Ω be optimal.2 Minimum-time problems . v ∗ (s)) 0 +λ∗ (s) f (t∗ (s). p∗ (t). on the other hand (8. (8. z ∗ (s). ∞) and let u∗ : [t0 .32) and (8. (8. x∗ (t). (t)) ≡ constant . u(t)). control constraint: u(·) ∈ U . not identically zero.

τ ) = so that the solution of (8.47) where α = (σ/m) > 0 and b = (1/m) > 0. Starting with an initial condition x(0) = x01 .108 CHAPTER 8.48) is p∗ (t) 1 p∗ (t) 2 or = e−α(t−τ ) − e−α(t−τ ) ) . We now study a simple example illustrating Theorem 2. 1 2 2 2 1 ∗ α p1 (0) 1 + eαt (− α p∗ (0) + p∗ (0)) . CONINUOUS-TIME OPTIMAL CONTROL Exercise 2: Prove Theorem 2. u ∈ R and u(t) constrained by |u(t)| ≤ 1. p∗ (t) ≡ p∗ (0) . x ˙ where m = mass. 1 1 and p∗ (t) = 2 The Hamiltonian H is given by H(x∗ (t). The control constraint set is Ω = [−1. p∗ (t). x2 = x we rewrite the particle dynamics as ˙ x1 (t) ˙ x2 (t) ˙ = 0 1 0 −α x1 (t) x2 (t) + 0 b u(t) . 1].48) such that (8.47) is 1 0 1 α (1 Φ(t. (8. 1 2 (8. (8. Now the transition matrix function of the homogeneous part of (8. x(0) = x02 we wish to ﬁnd an admissible control which brings the ˙ particle to the state x = 0. Example 1: The motion of a particle is described by m¨(t) + σ x(t) = u(t) . By Theorem 2 there exists a non-zero solution p∗ (·) of p∗ (t) ˙1 p∗ (t) ˙2 =− 0 0 1 −α p∗ (t) 1 p∗ (t) 2 (8. Suppose that u∗ (·) is optimal and x∗ (·) is the corresponding trajectory. ˙ Solution: Taking x1 = x. 1 1 αt α (1 − e ) 0 eαt p∗ (0) 1 p∗ (0) 2 . u = applied force. For simplicity we suppose that x ∈ R.44). and (8.49) .46) hold. σ = coefﬁcient of friction. x = 0 in minimum time. v) = (p∗ (t) − αp∗ (t))x∗ (t) + bp∗ (t)v 1 2 2 2 = eαt (p∗ (0) − αp∗ (0))x∗ (t) + pb∗ (t)v . and x = position of the particle.45).

50). from (8.51) We now proceed to analyze the consequences of (8. First of all since p∗ (t) ≡ 1 can have three qualitatively different forms.49) we see that p∗ (t) must be a strictly 1 2 2 monotonically increasing function so that from (8. 2 ∗ (t) > 0 for t > t. 2 ∗ −1 if p∗ (t) < 0.8. we must 1 2 2 1 have in this case p∗ (0) = 0. 2 109 (8.50) Furthermore. . since the right-hand side of (8. −p∗ (0) + αp∗ (0) > 0: Evidently then. −p∗ (0) + αp∗ (0) < 0 : Evidently u∗ (·) can behave in one of two ways: 1 2 either u∗ (t) = or u∗ (t) ≡ −1 and p∗ (t) < 0 for all t.50) u∗ (·) can behave in one of two ways: either ˆ ˆ −1 for t < t and p∗ (t) < 0 for t < t. VARIABLE FINAL TIME so that from the maximum principle we can immediately conclude that   +1 if p∗ (t) > 0. Also since p∗ (t) ≡ 0. Case 1. −p∗ (0) + αp∗ (0) = 0 : In this case p∗ (t) ≡ (1/α)p∗ (0). 2 ˆ and p∗ (t) < 0 for t > t. Case 3. ˆ −1 for t > t 2 >0.49) and (8.3. 1 2 2 2 p∗ (0). ˆ ˆ +1 for t > t and p2 u∗ (t) = or u∗ (t) ≡ +1 and p∗ (t) > 0 for all t. 2 Case 2. p∗ (·) 1 2 (8.47) does not depend on t explicitly we must also have eαt (p∗ (0) − αp∗ (0))x∗ (t) + bp∗ (t)u∗ (t) ≡ constant. u (t) = 2  ? if p∗ (t) = 0 . Hence u∗ (·) we can behave in one of two ways: 1 either u∗ (t) ≡ +1 and p∗ (t) ≡ 2 or u∗ (t) ≡ −1 and p∗ (t) ≡ 2 1 ∗ α p1 (0) 1 ∗ α p1 (0) ˆ ˆ +1 for t < t and p∗ (t) > 0 for t < t. <0.

53) (8. (8.110 CHAPTER 8.53) forward in time and check if (8. x20 = x20 also satisﬁes the ﬁnal condition x1 (t∗ ) = 0. Suppose we choose p∗ (0) such that −p∗ (0) = αp∗ (0) = 0 and p∗ (0) > 0.52) for some t∗ > 0. the optimal control u∗ is always equal to +1 or -1 and it can switch at most once between these two values. ξ2 (t) = b α (1 − eαt ) . This gives b ξ1 (t) = α (−t + eαt −1 α ) . ˙ 1 1 2 with initial condition x1 (0) = x10 . Let us follow this procedure. if p∗ (0) is such that −p∗ (0) + αp∗ (0) = 0 and p∗ (0) < 0. One way is to guess at the value of p∗ (0) and then integrate (8. An alternative is to guess at the value of p∗ (0) and then integrate (8. and (8. f f (8.54). and then t∗ is the minimum time. On the other hand. The latter approach is more advantageous because we know that any trajectory obtained by this procedure is optimal for initial conditions which lie on the trajectory. .52) and (8. If (8. The optimal control is given by u∗ (t) = sgn p∗ (t) 2 1 1 = sgn [ α p∗ (0) + eαt (− α p∗ (0) + p∗ (0))] . p∗ (0) such that the solution of the 1 2 differential equation x = x2 ˙ 1 1 x2 = −αx2 + b sgn[ α p∗ (0) + eαt (− α p∗ (0) + p∗ (0))] .52) and (8. f f There are at least two ways of solving the two-point boundary value problem (8. which is the curve OA in Figure 8.54) backward in time and check of (8.54) is satisﬁed.3.53).52) and (8. ξ2 (t) = − α (1 − eαt ) . which is the curve OB. then u∗ (t) ≡ −1 1 2 2 and we get b ξ1 (t) = − α (−t + eαt −1 α ) b .52).53) is satisﬁed. with ξ1 (0) − ξ2 (0) = 0 . x2 (t∗ ) = 0 . 1 1 2 Thus the search for the optimal control reduces to ﬁnding p∗ (0).54) (8.54) backward in time give us a trajectory ξ(t) where u ˙ ˙ ξ1 (t) = −ξ2 (t) ˙ ξ2 (t) = αξ2 (t) − b . Then we must have 1 2 2 ∗ (t) ≡ 1. CONINUOUS-TIME OPTIMAL CONTROL Thus. Integrating (8.54) is not satisﬁed then modify p∗ (0) and repeat.

Next suppose p∗ (0) is such that −p∗ (0) + αp∗ (0) > 0. This give us the curve OCD. Then [(1/α)p∗ (0) + 1 2 2 1 ˆ + p∗ (0))] will have a negative value for t ∈ (0.8. and p∗ (0) < 0. if we integrate (8.52) and (8. We see then that the optimal control u∗ (·) has the following characterizing properties: u∗ (t) = 1 if x∗ (t) is above BOA or on OA −1 if x∗ (t) is below BOA or on OB .3. . then u∗ (t) = 1 for t < t and u∗ (t) = −1 for t > t. Hence we can synthesize the optimal control in feedback from: u∗ (t) = ψ(x∗ (t)) where the B u∗ ≡ −1 x2 u∗ ≡ 1 u∗ ≡ −1 O x1 u∗ ≡ 1 A Figure 8.52). and we get the 2 2 curve OEF .3: Backward integration of (8. ∞).54) backwards in time we get trajectory ξ(t) where eαt (−(1/α)p∗ (0) 1 ˙ ξ(t) = −ξ2 (t) ˙ ξ2 (t) = αξ2 (t)+ ˆ −b for t < t ˆ . Hence. b for t > t with ξ1 (0) = 0. and p∗ (0) < 0.54). Finally if p∗ (0) is such that −p∗ (0) + 1 ˆ ˆ αp∗ (0) < 0. ξ2 (0) = 0. VARIABLE FINAL TIME u∗ ≡ −1 B C D u∗ ≡ 1 O E 111 ξ1 ξ2 u∗ ≡ 1 A F u∗ ≡ −1 Figure 8. (8. t) and a positive value for t ∈ 2 ˆ (t.4: Optimal trajectories of Example 1.

x2 ) = 1 if (x1 . T ] → Rn . positive deﬁnite matrix.57) (8. Gf is a given f × n matrix. v) = − 1 p∗ [x∗ (t) P (t)x∗ (t) + v Q(t)v] ˜ 2 0 +p∗ (t) [A(t)x∗ (t) + B(t)v] so that the optimal control u∗ (t) must maximize − 1 p∗ v Q(t)v + p∗ (t) B(t)v for v ∈ Rp .56) (8. ˙ 0 and p∗ (t)⊥T f (x∗ (t)) = {ξ|Gf ξ = 0} .112 CHAPTER 8. so that we must search for a number p∗ ≥ 0 and a function 0 ∗ : [0.60) 1 −1 p∗ Q (t)B 0 (8. u(·) piecewise continuous.55): T Minimize 0 1 [x (t)P (t)x(t) + u (t)Q(t)u(t)]dt 2 (8. CONINUOUS-TIME OPTIMAL CONTROL function ψ : R2 → {1. consider the optimal control problem (8.4) ψ(x1 . (8. The Hamiltonian function is H(t.55) subject to dynamics: x(t) = A(t)x(t) + B(t)u(t). ˙ initial condition: x(0) = x0 . and x0 ∈ Rn . such that p p∗ (t) = −p∗ (−P (t)x∗ (t)) − A (t)p∗ (t) . not both zero. Quadratic Cost An important class of problems which arise in practice is the case when the dynamics are linear and the objective function is quadratic. x2 ) is below BOA or on OB . ﬁnal condition: Gf x(t) = bf . We apply Theorem 1 of Section 2. this will imply 0 u∗ (t) = whereas if p∗ = 0. bf ∈ R f are given vectors. x2 ) is above BOA or on OA −1 if (x1 .58) (t)p∗ (t) .58) cannot have a maximum.4 Linear System.56) we assume that P (t) is an n × n symmetric. x∗ (t). (8. then we must have 0 p∗ (t) B(t) ≡ 0 because otherwise (8. In (8. T is a ﬁxed ﬁnal time.59) . p∗ (t). control constraint: u(t) ∈ Rp . −1} is given by (see Figure 8. Speciﬁcally. 8. positive semi-deﬁnite matrix whereas Q(t) is a p × p symmetric. 2 0 If p∗ > 0. 0 ≤ t ≤ T .

T ]. because if p∗ = 0 then from (8.56) 0 0 we can see that p∗ (t) = (Φ(T.61) Next we claim that if the system is controllable then p∗ = 0. and hence the optimal control is 0 given by (8. We are faced with the so-called singular case (because we are in trouble–not because the situation is rare). 0 ≤ τ ≤ T . v) is independent of v for values of t lying in a non-zero interval. THE SINGULAR CASE 113 We make the following assumption about the system dynamics. For further details regarding the solution of this boundary value problem and for related topics see (See and Markus [1967]). but then from (8. under the controllability assumption. (p∗ (t)/p∗ )) will satisfy all the necessary ˆ 0 0 conditions so that we can assume that p∗ = 1. implies ξ = 0 . p∗ > 0.59). control constraint: s(t) ∈ [0. s(·) piecewise continuous. t)B(t) = 0 .60) (p∗ (t)) Φ(T. The problem can be summarized as follows: T Maximize subject to ˙ dynamics: k(t) = s(t)f (k(t)) − µk(t) . The optimal trajectory and the optimal control is 0 obtained by solving the following two-point boundary value problem: x∗ (t) = A(t)x∗ (t) + B(t)Q−1 (t)B (t)p∗ (t) ˙ p(t) = P (t)x∗ (t) − A (t)p∗ (t) ˙ x∗ (0) = x0 . Gf x∗ (T ) = bf .5 The Singular Case In applying the necessary conditions derived in this chapter it sometimes happens that H(t. 0 ≤ t ≤ T initial constraint: k(0) = k0 . (8.) Let Φ(t. 8. In such cases the maximum principle does not help in selecting the optimal value of the control. then we must have p∗ (t) ≡ 0 which is a ˜ 0 contradiction. x∗ (t). p∗ (t). p∗ (T )⊥T f (x∗ (T )) . 1]. τ ) be the transition matrix function of the homogeneous linear differential equation x(t) = ˙ A(t)x(t). Assumption: The control system x(t) = A(t)x(t) + B(t)u(t) is controllable over the interval ˙ [0.8. Then the controllability assumption is equivalent to the statement that for any ξ ∈ Rn ξ Φ(t.61) we get p∗ (T ) = 0. Thus. Now if p∗ > 0 it is trivial that p∗ (t) = (1. We illustrate this by analyzing Example 4 of Chapter 1. 0 ≤ t ≤ T . (See (Desoer [1970]) for a deﬁnition of controllability and for the properties we use below. Hence if p∗ = 0. τ )B(τ ) = 0 . ﬁnal constraint: k(t) ∈ R . 0 c(t)dt = T 0 (1 − s(t))f (k(t))dt .5. t)) p∗ (T ) and hence from (8.

p∗ (t) > 1.65) (8. Assumption (8.6. Then by Theorem 1 of Section 2.67) The behavior of the solutions of (8. T ] → [0. fkk (K) < 0 for all k . if p∗ = 0 then from (8.62) k→0 (8. 1] at s∗ (t). CONINUOUS-TIME OPTIMAL CONTROL We make the following assumptions regarding the production function f : fk (k) > 0. Such solutions exist and are unique by virtue of the assumptions (8.68) (8.) < < < > .5. and fk (k) − µ > 0 according as k > kG whereas f (k) − µk < 0 according as k > kM . (8.62) says that the marginal product of capital is positive and this marginal product decreases with increasing capital. ˙ (8. be the corresponding trajectory of the capital-to-labor ratio.62) and (8. we note from (8.63) is mainly for technical convenience and can be dispensed with without difﬁculty. such that 0 p∗ (t) = −p∗ (1 − s∗ (t))fk (k∗ (t)) − p∗ (t)[s∗ (t)fk (k∗ (t)) − µ] ˙ 0 with the ﬁnal condition p∗ (T ) = 0 .63) Assumption (8. (k. First of all. p)−. k∗ (t). 0 ≤ t ≤ T . t)−planes in Figure 8. there exist a number p∗ ≥ 0. ˙ The maximum principle says that H(t.114 CHAPTER 8. Hence we must have p∗ > 0 and then by replacing (p∗ . p∗ ) we 0 0 0 0 can assume without losing generality that p∗ = 1. (8. p∗ (t).62) that kG < kM . p∗ ) by (1/p∗ )(p∗ .66) (8. not both identically zero. Here kG .63). lim fk (k) = ∞ . 1] is an optimal savings policy and let k∗ (t). which immediately implies that   1 if p∗ (t) > 1 ∗ s (t) = 0 if p∗ (t) < 1  ? if p∗ (t) = 1 We analyze separately the three cases above.64) and (8. s∗ (t) = 1 : Then the dynamic equations become ˙ k∗ (t) = f (k∗ (t)) − µk∗ (t) .64) and the maximum principle holds. so that (8. (See Figure 8.64) simpliﬁes to 0 p∗ (t) = −1(1 − s∗ (t))fk (k∗ (t)) − p∗ (t)[s∗ (t)fk (k∗ (t)) − µ] . Now suppose that s∗ : [0. t)− and (p. kH are the solutions of fk (kG ) − µ = 0 and f (kM ) − µk = 0. Futhermore.68) is depicted in the (k. and a function p∗ : [0. p∗ (t) = −p∗ (t)[fk (k∗ (t)) − µ] . T ] → R. s) = (1 − s)f (k∗ (t)) + p∗ (t)[sf (k∗ (t)) − µk∗ (t)] is maximized over s ∈ [0.65) we must also 0 have p∗ (t) ≡ 0. Case 1.

p∗ (t) = −fk (k∗ (t)) + µp∗ (t) . ˙ In turn then we must have k∗ (t) = 0 for t ∈ I so that s∗ (t)f (kG ) − µKG = 0 for t ∈ I . and hence. p∗ (t) < 1.7. ˙ giving rise to the behavior illustrated in Figure 8. so −fk (k∗ (t)) + µ = 0 for t ∈ I .) Evidently if p∗ (t) = 1 only for a ﬁnite set of times t then we do not have to worry about this case. p∗ (t) = 1. kG s∗ (t) = µ f (kG ) for t ∈ I . Case 2. We face the singular case only if p∗ (t) = 1 for t ∈ I. where I is a non-zero interval.5: Illustration for Case 1.66) ˙ we get −(1 − s∗ (t))fk (k∗ (t)) − [s∗ (t)fk (k∗ (t)) − µ] = 0 for t ∈ I .70) . or k∗ (t) = kG for t ∈ I .69) (8.5. s∗ (t) = 0: Then the dynamic equations are ˙ k∗ (t) = −µk∗ (t) . But then we have p∗ (t) = 0 for t ∈ I so that from (8. THE SINGULAR CASE 115 p l fk > µ fk < µ k kM f > µk kG p kM f < µk k t l t Figure 8. (8. s∗ (t) =?: (Possibly singular case. Case 3.8.

ˆ Exercise 1: A capital-to-labor ratio k is said to be sustainable if there exists s ∈ [0. In particular we must have k < k . Thus in the singular case the optimal solution is characterized by (8. k kG kM Figure 8. The reason for this term is contained in the following exercise. 1] such that ˆ ˆ − µk = 0. or (B) there exists t2 ∈ (0.9. We then have three possibilities depending on the value of k∗ (t2 ): (Bi) k∗ (t2 ) < kG : then p∗ (t2 ) < 0 so that p∗ (t) > 1 for t < t2 and we are in Case 1 so that ˙ ∗ (t) = 1 for t < t .62). First of all. s∗ (t) > 1 if k0 < kG . For t < t1 we either have p∗ (t) > 1. We face two possibilities: Either (A) p∗ (t) < 1 for all t < [0. t2 ).69) and (8.70). t2 ) so that p∗ (t) = 1. k∗ (t) = k0 e−µt . and s∗ (t) = µ(kG /f (kG )) for t ∈ (t1 . .65) we know that for t close to T.63). We can now assemble separate cases to obtain the optimal control. T ] and then s∗ (t) = 0. The capital-to-labor ratio kG is called the golden mean and the singular solution is called the golden path. T ) such that p∗ (t3 ) = 1. from the ﬁnal condition (8. CONINUOUS-TIME OPTIMAL CONTROL µk line of slope µ f (k) . T ) such that p∗ (t2 ) = 1 and p∗ (t) < 1 for t2 < t ≤ T .116 f CHAPTER 8. s∗ (t) = 0 if k > kG . (Biii) k∗ (t2 ) − kG : then we can have a singular arc in some interval (t1 . or we have p∗ (t) < 1. p∗ (t) < 1 so that we are in Case 2. k∗ (t) = kG . s 2 0 G (Bii) k∗ (t2 ) > kG : then p∗ (2 ) > 0 but then p∗ (t2 + ε) > 1 for ε > 0 sufﬁciently small and since ˙ p∗ (T ) = 0 there must exist t3 ∈ (t2 . Show that kG is the unique sustainable capital-to-labor ratio which maximizes ˆ sf (k) ˆ sustainable consumption (1 − s)f (k). (8.8. for 0 ≤ t ≤ T .6: Illustration for assumptions (8. This contradicts the deﬁnition of t2 so that this possibility cannot arise. as in Figure 8. The various possibilities are illustrated in Figure 8.

whereas for a discussion of the singular case consult (Kelley. and (Balakrishnan and Neustadt [1964]). BIBLIOGRAPHICAL REMARKS p k 117 l k kG p t l t Figure 8.7: Illustration for Case 2. et al. On the one hand these include extensions to inﬁnite-dimensional state spaces and on the other hand they allow for constraints on the state more general than merely initial and ﬁnal constraints. There is no single source of computational methods for optimal control problems. For an applications-oriented treatment of this subject the reader is referred to (Athans and Falb [1966]) and (Bryson and Ho [1969]).6. [1968]). That book contains many extensions and many examples and it is still an important source. treatment see (Neustadt [1969])..8. Several important generalizations of the maximum principle have appeared. but mathematically difﬁcult. 8. For a less rigorous treatment of state-space constraints see (Jacobson. However. (McReynolds [1966]). (Kelley [1962]).6 Bibliographical Remarks The results presented in this chapter appeared in English in full detail for the ﬁrst time in 1962 in the book by Pontryagin. the derivation of the maximum principle given in the book by Lee and Markus is more satisfactory. . et al. [1971]). For a uniﬁed. For applications of the maximum principle to optimal economic growth see (Shell [1967]). Among the many useful techniques which have been proposed see (Lasdon. also consult (Jacobson and Mayne [1970]). et al. and (Polak [1971]). cited earlier. et al. [1967])..

8: Case 3.118 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL p k 1 . The singular case. . k kG p kG t 1 t Figure 8.

9: The optimal solution of example. BIBLIOGRAPHICAL REMARKS 119 p∗ 1 t s∗ 1 t k∗ T T p∗ 1 t s∗ 1 t2 T t k∗ t2 T t Case (A) p∗ T p∗ Case (Bi) t2 T t . . . .8. t Case (Biii) . t1 . t t2 T s∗ 1 s∗ µkG f (kG ) t k∗ kG k0 t k∗ .6. t t1 t2 T . t Figure 8. .

120 CHAPTER 8. CONINUOUS-TIME OPTIMAL CONTROL .

·) : X × U → X are ﬁxed functions. u(i)) + Φ(x(N )) (9. In (9. However. . i = 0. In the ﬁrst section we develop the main recursion equation of DP for discrete-time problems.1 Discrete-time DP We consider a problem formulation similar to that of Chapter VI. the state x(i) and the control u(i) belong to arbitrary sets X and U respectively. 1. leads to optimality conditions which are sufﬁcient. . ·. The only disadvantage (which unfortunately often rules out its use) of DP is that it can easily give rise to enormous computational requirements. Φ : X → R. 121 . N − 1 . . ·) : X × U → R. . for notational convenience we neglect ﬁnal conditions and state-space constraints.Chapter 9 Dynamic programing SEQUENTIAL DECISION PROBLEMS: DYNAMIC PROGRAMMING FORMULATION The sequential decision problems discussed in the last three Chapters were analyzed by variational methods. Dynamic programming (DP is a technique which compares the optimal decision with all the other decisions. N −1 Maximize i=0 f0 (i. x(i). The second section deals with the continuous-time problem. u(i)) . ·. is that DP permits very general problem formulations which do not require differentiability or convexity conditions or even the restriction to a ﬁnite-dimensional state space. . besides the fact that it give sufﬁciency conditions. i = 0. The Ωi are ﬁxed subsets of U . 9. Some general remarks and bibliographical references are collected in the ﬁnal section. . N − 1 . 1. The main advantage of DP. x0 ∈ X is ﬁxed.. Finally f0 (i.1) subject to dynamics: x(i + 1) = f (i. therefore. X and U may be ﬁnite sets. This global comparison. initial condition: x(0) = x0 . or even inﬁnitedimensional spaces. control constraint: u(i) ∈ Ωi . .e. f (i. i.1). or ﬁnite-dimensional vector spaces (as in the previous chapters). x(i). the necessary conditions for optimality were obtained by comparing the optimal decision with decisions in a small neighborhood of the optimum. .

u(i)) + Φ(ˆ(N )) ˆ ˆ x i= N −1 (9. i = k. Proof: Suppose not. ˆ and the corresponding trajectory. . N − 1.2)k. for each x ∈ X and k between ) and N − 1. . Lemma 1: Suppose u∗ (k). Then for any . k + 1. . in which the system starts in state x0 at time 0. . N − 1 . x∗ (k + 1). such that ˆ ˆ ˆ N −1 f0 (i. i = k. . u(i)) + Φ(x(N )) . . x( + 1). . ˆ The value of the objective function corresponding to this control for the problem (9. . . ∗ ∗ ∗ > i= But then consider the control u(k). x(i). . is x(k). . . . . N − 1 . x(i). . . .2)k. u∗ (i)) + Φ(x∗ (N )) . x(N ) where ˜ ˜ x(i) = ˜ x∗ (i) . . . . u (i)) + Φ(x (N )) . . k + 1. . . starting in state x at time k.x . u(N − 1) with ˜ ˜ u(i) ˜ u∗ (i) . ·. with corresponding ˆ ˆ ˆ trajectory x( ) = x∗ ( ). . .2) .122 CHAPTER 9. u(i)) + Φ(ˆ(N )) ˆ ˆ x i= i=k N −1 f0 (i. . . . . .3) f0 (i.2) subject to dynamics: x(i + 1) = f (i. . . consider the following problem: N −1 Maximize i=k f0 (i.x∗ ( ) . and let x∗ (k) = x. Then there exists a control u( ). initial condition: x(k) = x. . k ≤ ≤ N − 1. − 1 u(i) . u∗ (i)) + N −1 f0 (i. objective function. . N . x (i). x∗ (i). x(i). u∗ (N − 1) is an optimal control for (9.x is N −1 f0 (i. (9. and control constraint as in (9. x(N ).1). . u∗ ( ). u∗ (N − 1) is an optimal control for (9. . . More precisely. i = + 1. x∗ (i). . . . i = . . u(i)) + Φ(˜(n)) ˜ ˜ x i=k −1 = > i=k f0 (i. x(i). . x(i) . . u(i)). x(i). .x to distinguish between different problems. . control constraint: u(i) ∈ Ωi . . Since the initial time k and initial state x are the only parameters in the problem above. . we will sometimes use the index (9. u(N − 1). . We begin with an elementary but crucial observation. into a family of optimal control problems with the same dynamics.1) but with different initial states and initial times. . i = k. x∗ (N ) be the corresponding optimal trajectory. u( + 1).2)k. i = k. DYNAMIC PROGRAMING The main idea underlying DP involves embedding the optimal control problem (9.

k ≤ ≤ N − 1 . .1. k ≤ ≤ N − 1. i.(end theorem) Corollary 1: Let u(k). so that u∗ (k). . f (x. Combining these two facts we get f0 (k. . N − 1 is an optimal feedback control. On the other hand. u∗ (N − 1) be an optimal control for (9. Then V ( . where . x∗ . Corollary 2: For k = 0. . x( )) ≤ f0 ( . u∗ (k)) .x(k+1) . . Then ψ(k.5) is equal to f0 (k. x)) + V (k + 1. u. ψ(k. x. . . u(i)) + Φ(x(N )) . u∗ (i)) + Φ(x∗ (N )) (9. x. x∗ ( )). . x.2)k . x. f (k. u(k))} . x(i). x. x(N ). u∗ (k))) ≥ f0 (k.. u(k))) . . . . . V (k. ))|u ∈ Ωk }. Theorem 1: Deﬁne V (N. ·). x. x. . (9. u) + V (k1 . .2)k+1.2)k. . x(i). and let x∗ (k) = x. contradicting the hypothesis. xu∗ (k)) + V (k + 1. . . x) satisﬁes the backward recursion equation V (k. 1. . .x and let x(k) = x. . u. ≥ i=k By Lemma 1 the left-hand side of (9. let ψ(k. (end theorem) From now on we assume that an optimal solution to (9.e. ·) : X → Ωk be such that f0 (k. and equality holds for all k ≤ ≤ N − 1 if and only if the control is optimal for (9. (k. u∗ (N − 1) deﬁned by u∗ ( ) = ψ( . let u∗ (k).x . x)) = Max{f0 (k. f ( . . x. Let V (k. x) = Max{f0 . for all u(k) ∈ Ωk . . f (k.2)k. We have N −1 i=k N −1 f0 (i. . . u( )) + V ( + 1. for any k. u∗ (N − 1) cannot be optimal for 9. x(N ) be the corresponding trajectory. . ·) by (V (N.2)k. u∗ (k)) + V (k + 1. . u) + V (k + 1. x. x the control u∗ (k). x( ). ψ(k. . x. We call V the (maximum) value function.2)k. by the deﬁnition of V we have N −1 f0 (i. u(i)) + Φ(x(N )) ≤ f0 (k.x . u(k)) i=k N +{ i=k+1 f0 (i. f (k. DISCRETE-TIME DP 123 by (9.2)k. . x. x. 0 ≤ k ≤ N − 1 . x) be the maximum value of (9. .5) f0 (i. . .9. u( )). . u(k)) + V (k + 1. . u(N − 1) be any control for the problem (9. (k)) + V (k + 1. u))|u ∈ Ωk } . u(N − 1) is optimal for (9. with equality if and only if u(k + 1).x . x∗ (N ) be the corresponding trajectory be x(k) = x. x. x) = Φ(x). f (k. f (k. x(i).4) Proof: Let x ∈ X. . and all x ∈ X. x( ). . k. k = 0.4).3). . u(i)) + Φ(x(N )) = f0 (k. . . f (k. x∗ (i). which is equivalent to (9.x exists for all 0 ≤ k ≤ N − 1. . . N − 1.

1. Note that this feedback control is optimum for all initial conditions.x .4) analytically. Remark: Theorem 1 and Corollary 2 are the main results of DP. x∗ ( )).6) In (9. x(t). 9. x(t).000. t + ∆] → Ω}. the DP formulation may necessitate a prohibitive amount of computation since we would have to compute and store the values of V and ψ for all k and x. Then we have to compute and store 10 × 20 values of V .2): Maximize 0 f f0 (t. How many units of each item should be placed in each knapsack so as to maximize total utility? Formulate this problem by DP. This “curse of dimensionality” seriously limits the applicability of DP to problems where we cannot solve (9.000. ∆ ≥ 0 . x) = Max{ t t+∆ f0 (τ. However. Then it is easy to see that V must satisfy V (t. He assumes that each person can take up to W pounds in his knapsack. let V (t. Φ : Rn → R is assumed differentiable and f0 . t0 ≤ t ≤ tf ˙ initial condition: x(0) = x0 . x∗ ( ). • Exercise 1: An instructor is preparing to lead his class for a long hike. f are assumed to satisfy the conditions stated in VIII.2 Continuous-time DP We consider a continuous-time version of (9. u(τ ))dτ (9. which is quite impractical for existing computers. The recursion equation (9.8) . and V (tf . u(t))dt + Φ(x(tf )) subject to dynamics: x(t) = f (t. k ≤ ≤ N − 1 . Then for N = 10. These numbers represent the relative utility of that item during the hike. unless we can ﬁnd a “closed-form” analytic solution to (9.4) allows us to compute the value function. For instance.7) +V (t + ∆. The instructor assigns a number Ui > 0 for each unit of item i. ψ( . suppose n = 10 and the state-space X is a ﬁnite set with 20 elements. for t0 ≤ t ≤ tf and x ∈ Rn . u ∈ Rp . tf ] starting in state x at time t. we have to compute and store 10x(20)n values of V . But now suppose X = Rn and we approximate each dimension of x by 20 values. As before.6). DYNAMIC PROGRAMING x∗ ( + 1) = f ( . x(τ ). and in evaluating the maximum in (9. which is a reasonable amount. t (9. There are N possible items to choose from.4). x ∈ Rn .000. and for n = 5 it is 32. (9. Ω ⊂ Rp . control constraint: u : [t0 . tf ] → Ω and u(·) piecewise continuous. Each unit of item i weighs wi pounds. For n = 3 this number is 80. x(t + ∆))|u : [t. u(t)) .4) we also obtain the optimum feedback control. x) be the maximum value of the objective function over the interval [t. x∗ (k) = x .124 CHAPTER 9.1. is optimal for (α)k. x) = Φ(x) .

x) + Max{f0 (t. x. u) + ∂V ∂x (t. Let u : [t. x. ψ(τ. ˆ Let x∗ (τ ) be the solution of x∗ (τ ) = f (τ. x(τ ) is the solution of x(τ ) = f (τ. x∗ (τ ))dτ dτ ∂V ∂V ∗ { (τ. (τ )). x∗ (τ ). CONTINUOUS-TIME DP In (9.6). u)|u ∈ Ω} = 0. t ≤ τ ≤ t + ∆ . u(τ )) . x(τ ). x.9) and the boundary condition (9. x) + ∂V f (t. u(τ )) . ψ(t. u(τ ))dτ + Φ(ˆ(tf )) .Bellman partial differentiable equation for the value function: ∂V ∂t (t.11) Note that the hypothesis concerning ψ guarantees a solution of (9. x∗ (tf )) − V (t. satisfying f0 (t.7). x∗ (t)) = f tf =∈t tf =− dV (τ. x. ˙ x(t) = x . x∗ (τ ) + x (τ )}dτ ˙ ∂τ ∂x t f (9. ∂x (9.2. x)) ∂x = Max{f0 (t.9) Theorem 1: Suppose there exists a differentiable function V : [t0 . x∗ (τ ))) . x∗ (τ ).14) F − 0(τ. tf ] and x ∈ Rn . x. tf ] × Rn → R which satisﬁes (9. and V is the value function. Let us suppose that V is differentiable in t and x. u) + ∂V f (t. x)) + ∂V f (t. Suppose there exists a function ψ : [t0 . ˆ ˆ ˆ x(t) = x . u)∆ + V (t. ˙ x∗ (τ ) = x . ψ(t. t . u∗ (τ ))dτ + Φ(x∗ (τ )) f0 (τ. t ≤ τ ≤ tf .8). x(τ ). u∗ (τ ))dτ . (9.9.13) ≤ To this end we note that V (tf . u)∆ ∂x + ∂V (t. To show that ψ is an optimal feedback control we must show that tf t tf t f0 (tτ.10) Then ψ is an optimal feedback control for the problem (9.12) · (9. (9. Let u∗ (τ ) = ψ(τ. tf ] → Ω be any piecewise continuous control and ˆ let x(τ ) be the solution of ˆ x (τ ) = f (τ. x. Then from (9. x)f (t. tf ] × Rn → Ω with ψ piecewise continuous in t and Lipschitz in x. x∗ . u)|u ∈ Ω} . x) = Max{f0 (t. x)∆ + o(∆)|u ∈ Ω}. x.7) we get V (t. x. x∗ (τ ). t ≤ τ ≤ tf .12). ˆ x (9. x∗ (τ ). ∂t 125 Dividing by ∆ > 0 and letting ∆ approach zero we get the Hamilton-Jacobi. t ≤ τ ≤ tf . Proof: Let t ∈ [t0 . ∆ > 0 .

For an important application of DP to computational considerations for optimal control problems see (Jacobson and Mayne [1970]). [] . x) and try a solution of the form V (t.126 using (9. x) = x R(t)x where R is unknown.15) using (9. x. • Exercise 1: Obtain the value function and the optimal feedback control for the linear regulatory problem: Minimize 1 x (T )P (T )x(t) + 2 1 2 T0 {x (t)P (t)x(t) +u (t)Q(t)u(t)}dt subject to dynamics: x(t) = A(t)x(t) + B(t)u(t) . From (9. Finally. x∗ (τ ). For an excellent introduction to this area of application see (Howard [1960]). the book of Bellman [1957] is still excellent reading.15). DYNAMIC PROGRAMING { ≤− t tf ∂V ∂V · x (τ )}dτ ˜ (τ.8) and the fact that x∗ (t) = x(t) = x we conclude that ˆ V (t. ˆ ˆ (9. It also follows that V is the maximum value function.9). u∗ (τ ))dτ . (9. Larson [1968] has developed computational techniques which greatly increase the range of applicability of DP where closed-form solutions are not available. 0 ≤ t ≤ T . V (tf . ˙ initial condition: x(0) = x0 . See (Bellman and Dreyfus [1952]) and (Wagner [1969]).9).] 9. where P (t) = P (t) is positive semi-deﬁnite. x(tf )) − V (t. and Q(t) = Q (t) is positive deﬁnite. (9. u∗ (τ )) f0 (τ. control constraint: u(t) ∈ Rp . (9.10). x) = Φ(x∗ (tf )) + ≥ Φ(ˆ(tf )) + x t tf t tf f0 (τ. The most elegant applications of DP are to various problems in operations research where one can obtain “closed-form” analytic solutions to be recursion equation for the value function.3 Miscellaneous Remarks There is vast literature dealing with the theory and applications of DP.13) is proved. In the case of sequential decision-making under uncertainties DP is about the only available general method.14). On the other hand. x(τ ). x(τ ). x(τ )) + ˆ ∂τ ∂x f0 (τ. [Hint: Obtain the partial differential equation satisﬁed by V (t. (t)) = ˆ ˆ t tf CHAPTER 9. u(τ ))dτ ˆ ˆ ♦ so that (9.

Dreyfus. [10] J. IEEE Trans. [6] R. (ed. 1969a. Hurwicz. on Systems Science and Cybernetics. Essays in Economics and Econometrics. Cullum.). [7] D. Director and R. Generalized Lagrange Multipliers in Dynamic Programming. Dantzig. Bruns. [15] S. CT-16(3). 127 . CT-16(3). McGraw-Hill. Polak. 1963. Desoer. Princeton University Press. Bellman and S. Bellman.D. and E. 1969. 1969b. Ho. [9] A. Notes for a Second Course on Linear Systems.E. Van Nostrand Reinhold. Rohrer. Applied Optimal Control.W. Decentralization and Computation in Resource in resource Allocation.E. Academic Press. 1960.W. [8] J. PhD thesis. Director and R. [2] M.A. Cannon. John Wiley. Applied Dynamic Programming. Blaisdell.E. CT-16(3). [12] C. On the design of resistance n-port networks by digital computer. 1964. Optimal Control. Theory of Optimal Control and Mathematical Programming. The function of operations research specilists in large urban schools.W. IEEE Trans.D.A.C.A. 1969c. Princeton University Press. University of California. 1970. IEEE Trans. [13] S. Computing Methods in Optimization Problems. McGraw-Hill. Athans and P. Rohrer. [5] R. [14] S. C. [4] K. Berkeley. 1971. Linear Programming and Extensions.J. Princeton University Press. IEEE Trans.W. 1957. [3] A.W.V. in Pfouts R. on Circuit Theory. 1954.L. 1970. Girshick. 1966. Bryson and Y.A.Bibliography [1] J. Banerjee. 1962. The generalized adjoint network and network sensitivities.A. on Circuit Theory.E. SSC-6(4). Blackwell and M. College of Engineering. Automated network design–the frequency-domain case. Rohrer. Arrow and L. Balakrishnan and L.E. on Circuit Theory. Dynamic Programming. Director and R. 1970. Theory of Games and Statistical Decisions. University of North Carolina Press. [11] G. Neustadt. Falb.

on Circuit Theory. New necessary condtions of optimality for problems with state-variable inequality constraints. MIT Press. [18] R. Greenberg and W. [19] Dowles Foundation Monograph. and Margolis.M. Linear Programming and Economic Analysis. . P. 18. Operations Research.A.M. Memo RM-6134-PR. [21] C. Gould. Jacobson. Public Expenditures and Policy Analysis. IEEE Trans. Introduction to Stochastic Control. 1967. The Rand Corporation. Rohrer. John Wiley. Rinehart. Math. 1970b. Mathematical Methods and Theory in Games. 1971.H. 1971. in Leitmann. Functions of Several Variables. 1970. 1971. [23] A. Operations Research. [22] D. [17] R. Apl. [32] S. Differential Dynamic Programming. Fleming.R. R. Addison-Wesley.I. Dynamic Programming and Markov Processes.Q. 1960. [31] D. Method of Gradients.J. A geometric duality theorem with economic applications. and J. in Haveman. [30] D. John Wiley.J. [28] R. 1959. Holt. 1970a. Primal resource directive approaches for optimizing nonlinear decomposable programs. [26] J. Production Theory and Indivisible Commodities. 1969. Geoffrion. The Economic Theory of Teams. 1970.P. Isaacs. to appear. Addison-Wesley. Dorfman and H. Review of Economic Studies. Duality in Nonlinear Programming: a Simpliﬁed Application-Oriented Treatment. [20] W.A. 1965.D.128 BIBLIOGRAPHY [16] R. Pierskalla. Dorfman. and Winston.H. Samuelson. (eds.). Surrogate mathematical programming. 1965. Academic Press. J. 18.Kushner. Howard. Karlin. Speyes.(ed. Analysis and Applications. SIAM J. [25] F.H. Jacobson and D. American Elsevier Publishing Co. to appear.. Lele.M. G.J. [27] H. Optimization Techniques. [24] A. Automated design of biasing circuits.M. Gale. Frank. Differential Games. Extensions of lagrange multipliers in nonlinear programming. CT-18(1). Markham Publishing Co. 1962. Jacoby. 1970. and R.). [29] R.L. volume 1. and Economics. McGraw-Hill. XXXIV(1). Programming. Dowell and R. 1969. Kelley. [33] H. 1958.A. Mayne. 17. M.Solow. 1971. Geoffrion. Princeton University Press.J. Math. J.H. A Model of Public Decisions Illustrated by a Water Pollution Policy Problem.

and Y. P. and H. Pontryagin.BIBLIOGRAPHY 129 [34] J.W. [38] R. Theory of Hierarchical. A general theory of extremals. Analysis and Applications. Lasdon. (eds. in Leitman. [49] L. Gamkrelidze.K. Mayer. 16.V. Louis. [36] T. Kendrick. [40] E. 1971.H. 1967.. .D. AC-12(1). R. [42] D. Macho. Foundation of Optimal Control Theory. Rao. [52] E. Nikaido. on Math. 1963. Topics in Optimization. Academic Press. [39] L. Applied Math. J. The conjugate gradient method for optimal control problems. Computational Methods in Optimization: A Uniﬁed Approach.. McGraw-Hill.). IEEE Trans.G. Nonlinear programming.S. The Mathematical Theory of Optimal Processes. 35(1). 1963.B. Water quality regulation with multiple polluters.W. Autom.E. (ed.R. S. Boltyanski. Kopp. Academic Press. 1970. 1967. D. J. [48] L. 1968. [50] H. Markus. Neustadt. Statistics and Probability. 1968. Waren. Math. Luenberger. Stochastic Optimal Linear Estimation and Control. and C. Larson. Polak. Math. H. 1968. [35] D.S. constraints.Mischenko. 1969.A. Second Berkeley Symp. St. J. Econometrica. 1970. Kelley. Saunders & Co. Games and Decisions.. Computer and System Sciences. Interscience. [37] H..W. 1967. In Proc. in Graves. Quasi-convex programming. Meditch. McGraw-Hill.G. Mangasarian. Control Conf.E.L. 1969. [45] J. 1967. University of California Press. 7.B. 1969. The Simplex Method for Local Separable Programming. The existence of optimal controls in the absence of convexity conditions.V. and Wolfe.C. 1957. John Wiley. Academic Press. G. and outcomes in optimal growth models. Miller. and E. Berkeley. In Proc. 19. 1968. R. Singular Extremals. Takahara. 1971 Jt. 1962. Lee and L. The successive sweep method and dynamic programming. [44] S. W. Recent Advance Programming.D. R. 1951. [43] O. Objectives. [53] L. John Wiley. 3(1).W. Washington U. Koopmans. [51] G. Mitter. R. Raiffa.). Analysis and Applications. Tucker. Kuhn and A. Neustadt. [47] C. Owen. Mesarovic. Wells. Multi-level Systems. and A.E. August 11-13 1971. Nonlinear Programming. Convex Structures and Economic Theory. State Increment Dynamic Programming.S.S.L. [46] M. American Elsevier Publishing Co.F. McGraw-Hill. McReynolds. Luce and H. [41] R.H. on Automatic Control. Game Theory. Siam J. Academic Press.

MIT Press. Topkis and A. Zangwill. On the convergence of some feasible directions algorithms for nonlinear programming.130 [54] R.T. [60] H. BIBLIOGRAPHY [55] M. Wonham. The economist’s approach to pollution and its control. Notes on Linear Programming. The simplex method for quadratic programming. [57] K. 27. Princeton University Press. 1969. on Control. on Control. Wolfe.M. 6(2). Science. [63] E. Shell. 1967. Wagner.J. John Wiley. 1968. 1954. Essays in the Theory of Optimal Economic Growth. The Foundation of Statistics.M. 1971. Prentice-Hall. Prentice-Hall. [58] R. 1967. . [62] W. [61] P. Principles of Operations Research. SIAM J. SIAM J. 1970. Van Nostrand Reinhold. Sakarovitch. 1959. Solow. 1971. [56] L. 5(2). [59] D. Nonlinear Programming: A Uniﬁed Approach. Veinnott Jr. Econometrica.M. On the seperation theorem of optimal control. 173(3996). Savage. 1969. Rockafeller.M. Convex Analysis.

121. 39 basic variable. 67 Convex function deﬁnition. 103 problem formulation. 5 Gradient. 63 Dynamic programming. 15 Discrete-time optimal control necessary condition. 37 properties. 125 problem formulation. 124 Lagrange multipliers. 101. 33. 105 continuous-time. 124 Epigraph. 101. 21 Linear programming. 35 problem formulation. 78. 103 discrete-time. 54 Basic feasible solution.LP optimality condition. 8 Hamilton-Jacobi-Bellman equation.Index Active constraint. 53 sufﬁcent conditions. 39 Certainty-equivalence principle. 125 ˜ Hamiltonian H. 42 Linear programming. 81 Minimum-time problem. 64 Farkas’ Lemma. 98 continuous-time. 80 Dual problem. 91. 78 problem formulation. 103 sufﬁcient condition. 99 ˜ Hamiltonian HH. 32 Feasible direction. 49 Game theory . 72 algorithm. 61 Knapsack problem. DP optimality conditions. 21. 34 Constraint qualiﬁcation deﬁnition. LP duality theorem. 37 Lagrangian function. 55 Continuous-time optimal control necessary condition. 35 Langrangian function. 33. 86. 80 Minimum fuel problem. 71 Feasible solution. 85 Adjoint equation augmented. 55 Convex set. 123 Discrete-time optimality control sufﬁcient condition. 54. 107 . 50 Adjoint Equation augmented. 5 Complementary slackness. 37. 125 Control of water quality. 80 Adjoint network. 58 131 Duality theorem. 77 sufﬁcient condition. 23 Afﬁne function. 34 Maximum principle continuous-time. 54 Langrangian multipliers. 37 Derivative. 123. 33. 8 Design of resistive network. 31 theory of the ﬁrm. 33.H. 91. 91 discrete-time. 61 Equilibrium of an economy. 101 Hypograph. 45. 101.

60 Supergradient. 103 Vertex. 50. 37 Phase I. 32 State-space constraint continuous-time problem. 11 sufﬁcient condition. 81. 2. 45. 41 Phase II. QP optimality condition. 37. 70 Wolfe algorithm. 49 suﬁcient condition. 77 Subgradient. 65 Separation theorem for convex sets. 38 Weak duality theorem. 17 sufﬁcient condition. 53 problem formulation. 5 Shadow prices. 39 Simplex algorithm. 80 Value function. 4 Optimization with equality constraints necessary condition. 33. 125 Optimization over open set necessary condition. 73 Separation theorem for stochastic control. 117 Optimal feedback control.132 example. 39 Nonlinear programming. 117 discrete-time problem. 33 Quadratic cost. 60 Supporting hyperplane. 58 Wolfe algorithm. 61. 70 Shadow-prices. 1 Optimal economic growth. 70 Primal problem. 21 Optimum tax. 13 Optimization under uncertainty. 108 Non-degeneracy condition. 54 Optimal decision. 123. NP duality theorem. 81. 91 discrete-time problem. 113 Slack variable. 70 problem formulation. 124 Regulator problem. 71 Recursion equation for dynamic programming. 50 Transversality condition continuous-time problem. 112 Quadratic programming. 123 Variable ﬁnal time. 71 INDEX . 112 Resource allocation problem. 63 necessary condition. 39 Singular case for control. 84 Tangent. 113.