Mathematical Methods of Economics

Joel Franklin
California Institute of Technology, Pasadena, California 91125

The American Mathematical Monthly, April 1983, Volume 90, Number 4, pp. 229–244.

W
hen Dr. Golomb and Dr. Bergquist asked me to give a talk on economics, my
first impulse was to try to get out of it.
“Sol,” I said, “I’m not an economist. You know that.”
“I know,” said Golomb.
“If you want an economist, I can get you one,” I said. “I know some excellent
economists.”
“No,” he said, “we want a mathematician to talk about the subject to other
mathematicians from their own point of view.”
That made sense, and I hit on this idea: I won’t try to tell you what mathematics has
done for economics. Instead, I’ll do the reverse: I’ll tell you some things economics has
done for mathematics. I’ll describe some mathematical discoveries that were motivated
by problems in economics, and I’ll suggest to you that some of the new mathematical
methods of economics might come into your own teaching and research.
One of these methods is called linear programming. I learned about it in 1958. I had just
come to Caltech as a junior faculty member associated with the computing center. The
director and I made a cross-country trip to survey the most important industrial uses of
computers. In New York, we visited the Mobil Oil Company, which had just put in a
multi-million-dollar computer system. We found out that Mobil had paid off this huge
investment in two weeks by doing linear programming.
Back at Caltech, Professor Alan Sweezy in economics and Professors Bill Corcoran and
Neil Pings in chemical engineering urged me to teach a course in linear programming.
When I told them I didn’t know linear programming, they said: Fine, Joel, learn it.
Seeing they meant business, I did study the subject and give the course. The students
loved it, and so did I. Perhaps you will have a similar experience.
One surprising thing I found was this: The mathematics was delightful. I knew it was
useful, but I hadn’t expected it to be beautiful. I was surprised to find that linear
programming wasn’t just business mathematics or engineering mathematics; it was the
general mathematics of linear inequalities. Later I found this mathematics coming into
some of my own special fields of research (statistics, numerical analysis, ill-posed
problems). Here again, you may have a similar experience.

The author is Professor of Applied Mathematics at the California Institute of Technology.

At Stanford University, in 1953, he received his Ph.D. in pure mathematics. His published
research pertains to numerical analysis, to ill-posed problems, to stochastic processes, and to
mathematical problems in engineering, in crystallography, in geophysics, and in cell biology.
At Caltech he has twice received the Associated Students Award for Excellence in Teaching.
He has written two textbooks for undergraduates in mathematics: Matrix Theory and Methods of
Mathematical Economics.
*Invited address to the Mathematical Association of America and to the Society for Industrial
and Applied Mathematics, November 14, 1981, Santa Barbara, California.
Linear programming is one of the many mathematical methods of economics. Here are a
few others: quadratic programming, geometric programming, general nonlinear
programming; fixed-point theorems—especially the Kakutani theorem; calculus of
variations, control theory, dynamics programming; theory of convex sets—especially
convex cones; probability, statistics, stochastic processes; finite structures (graph theory,
lattice theory); matrix theory; calculus, ordinary differential equations; and special
topics like game theory and Arrow’s theory of rational preference orderings.
Plato said mathematics is the essence of reality; Willard Gibbs said mathematics is the
language of science. If they are right, we shouldn’t be surprised to find uses for any
branch of mathematics in any science. Every branch of mathematics may have some use
in the science of economics. Here are two bizarre examples:
Have you heard of nonstandard analysis? I’ve heard of it, but know next to nothing
about it. Nevertheless, on November 10, 1981, I heard Yale Economics Professor Donald
J. Brown give a colloquium on the nonstandard analysis of hyper-finite economies (see
[4] and [20]).
You have heard of Bourbaki; so have I. I always thought that stuff would never be good
for anything. Nevertheless, Bourbaki ultrafilters appear in a paper in the Journal of
Economic Theory [17]. The authors, A. Kirman and D. Sondermann, use ultrafilters to
generalize Kenneth Arrow’s fundamental theorem of welfare economics [1].
Mathematics appears in all parts of economics, especially in mathematical economics
and in econometrics. Mathematical economics is like mathematical physics: it is
theoretical, nonempirical, sometimes speculative. For instance, Alfred Marshall
hypothesized the existence of certain curves (supply and demand schedules) whose
intersections determine commodity prices. Very pretty, but he didn’t show how to
measure or predict numerical values for specific supply-demand schedules.
In general, measurement and prediction belong to econometrics. As you would expect,
econometrics uses a lot of mathematical statistics, probability theory, and numerical
analysis. A Nobel prize was given in 1980 to Lawrence Klein for his work in building
econometric models.
In 1969 the first Nobel prize in economics was given to Ragnar Frisch and Jan
Tinbergen “for having developed and applied dynamic models for the analysis of
economic processes”; in other words, the prize was given for mathematics applied to
economics. Later, I’ll show you a list of all the Nobel prizes in economics, and you’ll
see that at least 7 of the 12 prizes given from 1969 through 1981 were given for work
that could be called applied mathematics. In fact, in 1975 a Nobel prize in economics
was given to Leonid Kantorovich, who is a mathematician.
In 1969 a spokesman for the Nobel foundation welcomed the new prize subject,
economics, as “the oldest of the arts, the youngest of the sciences.” It might be fair to
say that economics became a science when it started making significant use of
mathematics. When was that? I’d say the nineteenth century.
In 1817 the stockbroker David Ricardo proved a theorem that establishes an astounding
principle of international economics. Ricardo proved mathematically that free trade is
(under certain assumptions) advantageous to consumers in all nations.

2
Alfred Marshall was another great nineteenth-century economist. Marshall started out to
be a mathematician; he was First Wrangler in mathematics at Cambridge. Although his
work is seldom explicitly mathematical, any mathematician reading it can sense its
mathematical core. Marshall was a teacher of John Maynard Keynes, whose work
contains plenty of explicit mathematics. But, at least to my taste, Marshall’s work shows
more mathematical insight.
As Gerard Debreu wrote in his Theory of Value [7], mathematical economics has
become increasingly geometric and qualitative. If we want precise numerical
information, we have to turn to econometrics. Whereas Marshall drew his supply-and-
demand curves in a nonnumerical, qualitative way, the econometrician would have the
hard problem of giving numerical values for these curves for specific commodities at
specific times.
An example of econometrics appears in an article [29] by mathematician Jacob
Schwartz. He used a Wharton econometric model for residential housing. You can see it
in Fig. 1. There you see a typical awful equation of econometrics; please don’t try to
understand it. I just want you to see what is looks like. It predicts the rate of investment
in residential housing as a function of various factors (the numerical subscripts refer to
time lags). The coefficients s58.26, 0.0249, etc.d come from a numerical curve fit to data
for 1948–1964; the model was published in 1967.
There is an old Chinese proverb: It is always difficult to predict—especially the future.
For that reason econometrics is difficult. The Wharton model of 1967 “predicts” housing
starts for 1948–1964—not for the future. In general, econometric models are not laws of
nature like f 5 ma or E 5 mc2; they are empirical studies whose predictive value
depends on the constancy of the underlying relationships.

1967 Wharton econometric model (for 1948–1964)

Ih 5 58.26 1 0.0249Y 2 45.52 1pp 2

h

r 23
1 1.433siL 2 isd23 1 0.0851sIhs d21

Ih 5 rate of investment s $109 d in residential housing per quarter (3 months)

Y 5 total disposable income
ph 5 average housing price
pr 5 average rental price
iL 5 long-term interest rate
is 5 short-term interest rate
I hs 5 rate of housing starts
Negative subscripts denote time lags.
FIG. 1.

What Do Economists Think of Mathematics? That question has had different

answers at different times. Now the answer would be overwhelmingly favorable, if not
unanimous. But not so in the old days. Adam Smith published his great book Wealth of
Nations in 1776. It is readable, fascinating, and important; but it contains almost no
mathematics. 3
I told you the great nineteenth-century economist Alfred Marshall had been First
Wrangler in mathematics at Cambridge. Later, he talked about the role mathematics
played in his work:
I had a growing feeling in the later years of my work at the subject that a good mathematical
theorem dealing with economic hypotheses was very unlikely to be good economics: and I
went more and more on the rules—(1) Use mathematics as a shorthand language, rather
than as an engine of inquiry. (2) Keep to them till you have done. (3) Translate into
English. (4) Then illustrate by examples that are important in real life. (5) Burn the
mathematics. (6) If you can’t succeed in 4, burn 3. This last I did often. —quoted in [31],
p. 307.
So Marshall practiced mathematics as a secret vice; he was a closet mathematician.
His most famous student was John Maynard Keynes. At Cambridge, Keynes took his
degree in mathematics*. In 1920 Keynes published his Treatise of Probability. Keynes’s
great books on economics contain many equations. By the time of Lord Keynes
mathematics was not a secret vice but a public virtue.
A living disciple of Keynes, Harvard Professor John Kenneth Galbraith, regards
mathematics with skepticism. One of Galbraith’s more entertaining books is called
Economics, Peace, and Laughter. Commenting on the models of mathematical
economics, he says this:
Moreover, the models so constructed, though of no practical value, serve a useful academic
function. The oldest problem in economic education is how to exclude the incompetent . . . .
The requirement that there be an ability to master difficult models, including ones for which
mathematical competence is required, is a highly useful screening device.
Not satisfied with this comment, Galbraith adds a dour footnote:
There can be no question, however, that prolonged commitment to mathematical exercises in
economics can be damaging. It leads to the atrophy of judgment and intuition . . . .
John Galbraith does not stand alone. He tells this story about Paul Samuelson, a superb
applied mathematician and winner of the Nobel Prize for work in mathematical
economics:
Professor Samuelson, in his presidential address to the American Economic Association
several years ago, noted that the three previous presidential addresses had been devoted to a
denunciation of mathematical economics and that the most trenchant had encouraged the
audience to standing applause.
Well! And skepticism about mathematics is not confined to this continent.
Galbraith says:
Once when I was in Russia on a visit to Soviet economists, I spent a long afternoon
attending a discussion on the use of mathematical models in plan formation. At the
conclusion an elderly scholar, who had also found it very heavy going, asked me rather
wistfully if I didn’t think there was still a “certain place” for the old-fashioned Marxian
formulation of the labor theory of value.

*While studying for the Tripos, Keynes wrote to his friend B. W. Swithinbank on 18 April 1905:
“I am soddening my brain, destroying my intellect, souring my disposition in a panic-stricken
attempt to acquire the rudiments of the Mathematics.” See R. F. Harrod [13], p. 130.
4
The old Russian scholar must have sighed when a Nobel prize in economics was given
to Leonid Kantorovich, a mathematician. Kantorovich got the prize for developing the
mathematical theory of linear programming and for applying it to the economic problem
of optimum allocation of resources. He would have gone a lot farther with linear
programming if he hadn’t run into trouble from the orthodox Marxians, who objected to
the use of the idea of prices. Dantzig tells the story in his book [6], p. 23.
Among the Nobel Laureates in economics, some, like Kantorovich, solved problems
in economics by inventing new mathematics; others made much use of known
mathematics. Look at the list of Nobel prizes in economics, Fig. 2. I’ve put asterisks
by seven of the twelve prize years to indicate work that is heavily mathematical.
Nobel Prizes in Economics
1969* Frisch, Ragnar and Tinbergen, Jan—“for having developed and applied dynamic
models for the analysis of economic processes.”
1970* Samuelson, Paul—“for the scientific work through which he has developed static and
dynamic economic theory and actively contributed to raising the level of analysis in
economic science.”
1971 Kuznets, Simon—“for his empirically founded interpretation of economic growth
which has led to new and deepened insight into the economic and social structure
and process of development.”
1972* Hicks, Sir John R. and Arrow, Kenneth J.—“for their pioneering contributions to
general economic equilibrium theory and welfare theory.”
1973 Leontief, Wassily—“for the development of the input-output method and for its
application to important economic problems.”
1974 Myrdal, Gunnar and Von Hayek, Friedrich August—“for their pioneering work in the
theory of money and economic fluctuations and for their penetrating analysis of the
interdependence of economic, social and institutional phenomena.”
1975* Kantorovich, Leonid and Koopmans, Tjalling—“for their contributions to the theory
of optimum allocation of resources.”
1976* Freidman, Milton—“for his achievements in the fields of consumption analysis,
monetary history and theory and for his demonstration of the complexity of
stabilization policy.”
1977 Ohlin, Bertil and Meade, James—“for their pathbreaking contributions to the theory
of international trade and international capital movements.”
1978 Simon, Herbert A.—“for his pioneering research into the decision-making process
within economic organizations.”
1979 Lewis, Arthur and Shultz, Theodore—for studies of human capital.
1980* Klein, Lawrence—for computer models designed to forecast economic changes.
1981* Tobin, James—for mathematical models of investment decisions.
* Asterisks indicate very mathematical work.
FIG. 2.

5
Seven out of twelve Nobel prizes—not a bad score for mathematics. And some of this
mathematics has freshness and charm. For example, let me show you a theorem that
won a Nobel prize: the Possibility Theorem of Kenneth Arrow.
In 1957 Kenneth Arrow published a little book called Social Choice and Individual
Values. He was thinking about a problem of welfare economics: Confronted by
numerous conflicting special interests, how should the government make decisions?
Use old-fashioned majority rule, you say. That’s the democratic way isn’t it? That’s the
rational way.
Let’s see. Suppose we have 3 alternatives: vanilla (V), chocolate (C), and strawberry
(S). And suppose we have 9 voters, each with his own individual values. For example,
one individual may like vanilla better than chocolate sV > Cd, and he may like
chocolate better than strawberry sC > Sd; then, by the way, he must like vanilla better
than strawberry sV > Sd if his individual values are rational. Another individual may
prefer strawberry to vanilla sS > Vd, vanilla to chocolate sV > Cd, and therefore
strawberry to chocolate sS > Cd. And so on.
If all of our nine voters have definite flavor preferences, the voters constitute 6 special-
interest groups, corresponding to the six ways of ranking 3 flavors. For example, we
might have the following tabulation:
Individual values Number of individuals
V > C > S 2
S > V > C 2
C > S > V 2
V > S > C 1
C > V > S 1
S > C > V 1
Now comes the general election. Here are the results:
V > C by a majority of 5 to 4
C > S by a majority of 5 to 4
and—what’s this?
S > V by a majority of 5 to 4.
But that’s crazy: V > C and C > S should imply V > S, not S > V. (This is an
example of Concordet’s paradox.)
No wonder Congress is confused. You see the problem. So did Arrow, and he wondered
if there was any way out.
There is one way out: Hitler’s way. Pick one individual, call him der Fuhrer, and do
what he says. Then all the government’s preferences can be nice and transitive, and too
bad for you if you don’t like it.
Is there any rational way to make social choices besides dictatorship? To this basic
question of welfare economics, Kenneth Arrow gave an astonishing answer: No.

6
ARROW’S THEOREM. Suppose we have a function that makes rational (transitive)
social choices as a function of rational individual values that rank (by preference or
indifference) three or more alternatives. Assume that the social-choice function has two
properties:
(i) If all individuals prefer alternative a to alternative b, then society shall prefer a
to b.
(ii) The social choice between any two alternatives a and b shall depend only on the
individual values between a and b (and should not depend on any third
alternative c).
Then Arrow’s theorem says there exists a dictator—a single individual whose
preferences become social choices.
In a minute I’ll write this theorem symbolically, in terms of matrices. But first I want to
explain the two assumptions. The first is a principle of unanimity: If everyone prefers
vanilla to chocolate, so should society. The second is a principle of relevance: Society’s
choice between vanilla and chocolate should depend on how people feel about vanilla
and chocolate, not on how they feel about strawberry.
If you wish, you can write Arrow’s theorem in terms of matrices. Let aij 5 1 if i is
preferred to j; let aij 5 21 if j is preferred to i; let aij 5 0 if neither is preferred to the
other. If there are m alternatives (flavors), then the numbers aij constitute an m 3 m
skew-symmetric matrix, A. In a rational preference ordering, if i is preferred to j, and if j
is preferred to k, then i must be preferred to k. For the matrix A this says: If aij 5 1 and
ajk 5 1, then aik 5 1. We shall also require aik 5 1 if aij 5 0 and ajk 5 1 or if aij 5 1
and ajk 5 0. If this is so, then we’ll call A a rational preference matrix.
EXAMPLE. Suppose we prefer flavor 3 to flavor 1 and flavor 2, which we like equally.
Then this is our rational preference matrix:

1 2
0 0 21
A5 0 0 21
1 1 0
EXAMPLE. Suppose we prefer flavor 1 to flavor 2, flavor 2 to flavor 3, and flavor 3 to
flavor 1. That is irrational, and so the preference matrix is irrational:

1 2
0 1 21
A5 21 0 1 .
1 21 0
Look: a12 5 1 and a23 5 1, but a13 Þ 1.
Individual values and social choice: Suppose there are n individuals and m alternatives.
The individual values are expressed by n rational preference matrices As1d,. . ., Asnd. A
social choice is a rational preference matrix A. We’re looking for a function F mapping
Pmn into Pm, where Pm is the set of m 3 m rational preference matrices and Pmn is the n-
fold Cartesian product:
A 5 FsAs1d,. . ., Asndd.

7
EXAMPLE. For majority rule, the function F is defined as follows:
aij 5 signfaijs1d 1 . . . 1 aijsndg si, j 5 1,. . .,md.
If m > 2, majority rule may give irrational social choices, as we saw in the example of
vanilla, chocolate, and strawberry. So this F takes values outside Pm; but this F does
satisfy the assumption of unanimity and relevance:
(1) aij 5 1 if aijskd 5 1;k 5 1,. . ., n
(2) aij 5 a function of aijs1d,. . ., aijsnd.
Arrow’s theorem now takes this form: Let F be a function mapping Pmn into Pm.
Suppose m > 2, and suppose the function F satisfies equations (1) and (2). Then there
exists an integer d such that aij 5 1 if aijsdd 5 1. (The integer d depends on F but not
on the matrices As1d,. . ., Asnd.)
By the way, there are no restrictions on the number of individuals, n. In marriage,
n 5 2. Then Arrow’s theorem says: Either the husband or the wife must be a dictator, or
there must be irrational choices. Experience seems to bear this out.
Arrow’s theorem talks about rational (transitive) preference orderings. This raises a
question in combinatoric analysis: How may rational preferences orderings of m
alternatives are there? The answer has appeared in [12]. For large m the number of
rational preference orderings behaves like s1y2dm!slog2d2m21.
The mathematics of Arrow’s theorem is very different from mathematics like linear
programming. Here we have a rather ordinary looking problem:
For i 5 1,. . ., m and j 5 1,. . ., n we are given the real numbers aij, bi, cj. We wish to
find numbers xj > 0 such that
n

o c x 5 minimum
j51
j j

over all solutions of the linear equations

n

oa x 5b
j51
ij j i si 5 1,. . ., md.

That is the canonical form of linear programming. In terms of matrices and vectors, it
looks like this:
Ax 5 b, x > 0, cTx 5 min.
The problem is interesting only if the linear system Ax 5 b has more than one solution,
so we usually suppose rank A 5 m < n. Then the crucial assumption is the sign
constraint x > 0 (all components of x must be nonnegative).
Kantorovich in Russa and Dantzig in the United States independently developed linear
programming to solve economic logistical problems. The history of their work appears
in Dantzig’s book [6].
The most famous early problem of linear programming, the diet problem, first appeared
in the Journal of Farm Economics [33]. The problem is to design a nutritionally
adequate diet at minimum cost. The author, George Stigler, won the 1982 Nobel Prize in
Economics.
8
Suppose aij is the amount of nutrient i in one unit of food j. (For instance, a37 might be
the amount of vitamin B1 in one gram of wheat bread.) Let bi be the minimum daily
requirement of nutrient i, and let cj be the cost of one unit of food j. Let xj be the
amount of food j in a daily diet. Then we require
n n

o
j51
aij xj > bisi 5 1,. . ., md, xj > 0 o c x 5 minimum.
j51
j j

This is a linear program in standard form. To put it in canonical form, we must replace
the m linear inequalities by equations. We do that by introducing m new unknowns
zi > 0:
n

o a x 2 z 5 b.
j51
ij j i i

The problem is now easy to solve by Dantzig’s simplex method.

Linear programming has many uses in industry and banking. In 1981, a good popular
article [2] appeared in Scientific American; I recommend its example on beer. An
introduction to the use of linear programming for the optimization of bank investment
portfolios appeared in the Monthly Review of the Federal Reserve Bank of Richmond
(see [3] and [10], p. 3). Banks and oil companies make a lot of money with linear
programming.
But you and I are mathematicians; money means nothing to us. So let us speak of
something more important—let’s talk about Chebyshev approximation.
Suppose we are given a system of real linear equations, Ax 5 b, and suppose the system
has no solution x. Typically, this occurs when we have more equations than unknowns. If
we have m equations in n unknowns, the error in equation i is a function of the vector x:
n
ei 5 oa x 2b
j51
ij j i si 5 1,. . ., md.

The problem of Chebyshev approximation is to find a vector x that minimizes the

maximum absolute error:
Minimize smax ei d.
x i
||
That is a beautiful and important problem of approximation theory. Many things were
known about Chebyshev approximation before 1959, but no one knew a good way to do
it. Then Edward Stiefel discovered how to do it by linear programming (see [32] and
[10], p. 8). Here’s how:
||
Define a new unknown: x0 5 max ei for i 5 1,. . ., m. Then we shall have the uniform
error bracket
n
2x0 < o a x 2b <x
j51
ij j i 0 si 5 1,. . ., md.

The problem of Chebyshev is to choose x0,. . ., xn so as to minimize the maximum

absolute error: Minimize x0.
That’s all there is to it—a finite number of linear inequalities in a finite number of
unknowns, with a linear form to be minimized. That is a linear program in general form.
9
It’s trivial to restate it in canonical form, and it’s routine to solve it numerically by the
simplex method.
The simplex method is perhaps the most important numerical method invented in the
twentieth century. Experience with enormous industrial problems shows that the simplex
method works fast. In problems with m equations in n unknowns, the computation time
seems to be proportional to n.
Why does the simplex method usually work so fast? No one knows, and this is one of
the great unsolved problems of numerical analysis. At first glance, the computation time
would seem to be proportional to the binomial coefficient s m d, which is the possible
n
number of basic solutions of Ax 5 b. For m , ny2, the binomial coefficient is almost as
big as 2n, and this suggests the computing time could grow exponentially with n. Indeed,
Victor Klee and George Minty [18] have constructed pathological cases for which that
happens. But it never seems to happen in practice.
A Russian mathematician named Khachian got around this problem by analyzing a quite
different algorithm [16]. Khachian proved that his algorithm has computing time
bounded by a constant, K, times n6—which becomes smaller than 2n. Khachian’s proof is
a triumph of theoretical computer science. But Khachian’s algorithm, in its present form,
has little practical value: the constant K is enormous and so is the computing time.
You can become famous by doing one of these two things: (1) show why the simplex
method usually works as well as it does; (2) show how Khachian’s method can be made
to work better than the simplex method in practice. [A persistent rumor says Stephen
Smale has done (1).]
Linear programming is important because it is the general mathematics of finite systems
of linear inequalities. Linear programming is more general than real linear algebra, for
this reason:
Any real linear equation oa x 5 b can be restated as a pair of linear inequalities:
i i

oa x < b and oa x > b.

i i i i

But the converse is false: You can’t restate a linear inequality as a finite number of
linear equations.
No mathematician doubts the importance of linear algebra. So linear programming must
also be important, and perhaps you will agree that linear programming should be part of
the basic undergraduate mathematics curriculum. Why should mathematics students
have to pick up their linear programming from economists and chemical engineers and
people like that? They should learn it from us, and they should learn it right.
Marshall Hall has a section on linear programming in his book Combinatoric Analysis.
There’s nothing odd about that; linear programming has many applications to
combinatorics. For instance, look at this problem:
We are given an n 3 n matrix of real numbers aij. We seek a permutation ji,. . ., jn that
maximizes the sum
s 5 a1j 1 a2j 1 . . . 1 anj .
1 2 n

10
This problem is called the optimal-assignment problem.
EXAMPLE. Suppose we’re given the matrix

1 2
7 2 6
3 9 1 .
8 4 5
The sum s has six possible values. The largest is
max s 5 a13 1 a22 1 a31 5 6 1 9 1 8 5 23,
achieved for the permutation s j1, j2, j3d 5 s3, 2, 1d.
In general, we could solve the problem by calculating all the n! possible values for s, but
that takes too long if n is large. A much faster algorithm is given by linear programming.
We define the unknowns xij as 1 if j 5 ji, or 0 if j Þ ji. Thus, xij will tell us which
component to pick from each row. For the preceding numerical example, we would have

1 2
0 0 1
sxijd 5 0 1 0 .
1 0 0
In general, the integer unknowns xij must satisfy the constraints
xi1 1 . . . 1 xin 5 1 si 5 1,. . ., nd
xij 1 . . . 1 xnj 5 1 sj 5 1,. . ., nd
xij > 0 si, j 5 1,. . ., nd.
Then we wish to maximize a linear form:

s5 oa x
i, j
ij ij 5 maximum.

This is a problem in linear programming. H. W. Kuhn [19] has shown that it can be
solved in Osn3d steps.
You are right if you object that linear programming provides the optimal real solution
xij, and these numbers might not be integers (we need all xij 5 0 or 1). But for the
optimal-assignment problem the optimal solution over the integers xij is also optimal
over the real numbers xij. That’s not obvious, but it’s easy to prove. In general, however,
linear programming over the integers is difficult. The optimal solution over integers is
usually not optimal over real numbers.
So much for combinatorics. Now let’s look at geometry. I’d like to show you how
quadratic programming solves a problem stated in 1857 by J. J. Sylvester [34]: “It is
required to find the least circle which shall contain a given set of points in the plane.”
Suppose the given points are a1,. . ., an. We’re looking for a circle with the unknown
center x and radius r. The given points are required to lie inside the circle:
iai 2 xi2 < r2 si 5 1,. . ., md.
Then we want to choose x and r so as to minimize r.

11
We can replace the m quadratic inequalities by linear inequalities as follows. Introduce
the unknown
1
x0 5 s r2 2 ixi2d.
2
Then the m inequalities become
x0 1 ai ? x > bi si 5 1,. . ., md,
where bi 5 12 iai i2. Then we want to minimize r2:
2x0 1 ixi2 5 minimum.
Sylvester’s problem now has this form: First we require m linear inequalities:
x0 1 ai1 x1 1 ai 2 x2 > bi si 5 1,. . ., md.
Then we want
2x0 1 x12 1 x22 5 minimum,
in which the quadratic terms constitute a positive definite form. This is a routine
problem of quadratic programming. It can be solved numerically by an ingenious variant
of the simplex method. This algorithm was discovered by a mathematician, Philip
Wolfe, but it was published in an economics journal, Econometrica [36].
Why in an economics journal? Because Wolfe’s paper extended the work of some
economists who were interested in the use of quadratic programming to make optimal
investment decisions. Wolfe’s mathematical discovery solved a problem in economics.
The theoretical basis of linear and nonlinear programming was published in 1902 by a
mathematician named Julius Farkas. He gave a long, cumbersome proof of the following
proposition, which you might call the alternative of linear inequalities (generalizing the
Fredholm alternative of linear equations):
THE FARKAS THEOREM. Let A be a given m 3 n real matrix, and let b be a given
vector with m real components. Then one, and only one, of the following alternatives
is true:
(i) the system Ax 5 b has a solution x > 0 (all components > 0d;
(ii) the system of inequalities y TA > 0 has a solution y satisfying y Tb < 0.
Indeed, both alternatives can’t be true, for then we could deduce
0 < s y TAdx 5 y TsAxd 5 y Tb < 0.
That’s easy; the hard part is to show that one of the alternatives must be true. A modern
straightforward proof of the Farkas theorem relies on the separating-plane theorem for
convex sets (see, e.g., [10], p. 56).
The Farkas alternative has many uses outside mathematical economics. I hope to
convince you that every mathematician should know the Farkas theorem and should
know how to use it. For example, let me show how to use the Farkas theorem to prove
the fundamental theorem of finite Markov processes.

12
THEOREM (Markov). Suppose pij > 0, and suppose
n

op
i51
ij 5 1 s j 5 1,. . ., nd.

Then there exist numbers xj > 0 satisfying

n

op x 5 x
j51
ij j i si 5 1,. . ., nd
n

ox
j51
j 5 1.

The proof of a special case of this theorem occupies several pages in Feller’s book on
probability ([8], pp. 428–432). The general case is usually proved by using the Perron-
Frobenius maximum principle for positive matrices or by using the Brouwer fixed-point
theorem. Instead, we can give an elementary proof using the Farkas theorem ([10], p.
58):
First, we state Markov’s assertion as one Farkas alternative:
(i) There exists a vector x > 0 satisfying the n 1 1 linear equations
n

os p
j51
ij 2 dijdxj 5 0 si 5 1,. . ., nd
n

ox
j51
j 5 1,

where dij is the Kronecker delta.

Second, we state the other Farkas alternative:
(ii) There exist numbers y1,. . ., yn, yn11 satisfying the inequalities
n

o y sp
i51
i ij 2 dijd 1 yn11 > 0 s j 5 1,. . ., nd

yn11 , 0.
Alternative (ii) implies the strict inequalities
n

oy p
i51
i ij > yj for all j.

But
n
max yi > oy p
i51
i ij

because we assumed pij > 0 and oi pij 5 1, so we find

max yi > yj for all j.
That is impossible, so alternative (ii) is false.
Now Farkas tells us that alternative (i) is true: Markov’s theorem is proved. That was
easy, wasn’t it?

13
Now let me tell you about the theory of games and economic behavior. A book with that
title was published in 1944 by the mathematician John von Neumann and the economist
Oskar Morgenstern [25]. Economists consider this book an epoch-making contribution
to economics.
Fine, you say, but what has it done for mathematics?
This book, along with von Neumann’s earlier work [24] on game theory, has given us
some stimulating problems and some important results. For example, look at this
theorem on matrices:
THEOREM (VON NEUMANN). Let A be a real m 3 n matrix. Let vectors x and y range
over the sets
m n

o
i51
xi 5 1, xi > 0; o y 5 1,
j51
j yj > 0.

Then
min max xTAy 5 max min xTAy.
y x x y
This theorem is no platitude. As a rule, mixed extrema are not equal, as the following
example shows. Suppose x and y range over the sets 0 < x < 1, 0 < y < 1. Then
1
min max sx 2 yd2
5 ,
y x 4
but
max
x
min
y
sx 2 yd2 5 0.
Von Neumann’s minimax theorem is the fundamental result in the theory of zero-sum
two-person games. But that’s not the point; the point is, it’s good mathematics. Von
Neumann proved the minimax theorem by using the Brouwer fixed-point theorem. His
proof is nonelementary and nonconstructive. Later, the mathematician George Dantzig
gave an elementary, constructive proof by using the dual simplex method of linear
programming.
Following von Neumann, mathematical economists make much use of the fixed-point
theorems. Their favorite seems to be the fixed-point theorem of Kakutani [15].
As a young mathematician at the Institute of Advanced Study, Shizuo Kakutani
discovered a generalization of the Brouwer fixed-point theorem. Kakutani’s work was
motivated by problems in economic game theory. His theorem has great mathematical
novelty. It speaks of point-to-set mappings:
THEOREM (Kakutani). Let X be a closed, bounded, convex set in Rn. For every point x in
X, let Fsxd equal a nonempty convex subset of X. Assume that the graph
Hx, y : y [ FsxdJ is closed.
Then some point in X satisfies x * [ F sx *d.
The image of each point x is a convex set Fsxd , X. The theorem says some point x* lies
in its image Fsx*d. Figure 3 illustrates this. Kakutani’s theorem is novel because it talks
about set-valued functions.
14
 FIG. 3.

If every set Fsxd contains just one point, the closed-graph assumption is equivalent to
the continuity of the function Fsxd and then Kakutani’s theorem reduces to the Brouwer
fixed-point theorem. Kakutani proved his theorem by using the Brouwer theorem.
A private survey indicates that 96% of all mathematicians can state the Brouwer fixed-
point theorem, but only 5% can prove it. Among mathematical economists, 95% can
state it, but only 2% can prove it (and these are all ex-topologists). This dangerous
situation will soon be remedied. Within the last two years, John Milnor [22] and C. A.
Rogers [27] have produced elementary proofs, using nothing more advanced than
calculus. These proofs are so easy that I can understand them [10], and certainly you can.
While 96% of mathematicians can state the Brouwer fixed-point theorem, only 7% can
state the Kakutani theorem. This situation is also dangerous, or, at least, wasteful. The
Kakutani theorem has many potential applications outside economics; these applications
should be made. Now that we can all understand the Brouwer theorem, we can also
understand the Kakutani theorem, so nothing can stop us.
In the application of Kakutani’s theorem to many-person game theory, the point x
denotes a collection of mixed strategies and the set-valued function denotes the sets of
optimal mixed strategies. The inclusion x [ Fsxd characterizes an equilibrium solution
of the game. The Kakutani theorem is thus the perfect tool for proving J. F. Nash’s
fundamental theorem [23] on n-person games.
Professor H. F. Bohnenblust once told me something about research. He had supervised
many successful Ph.D. thesis projects—and a few unsuccessful ones. He said this: The
unsuccessful projects start with some famous old problem (prove the Riemann
hypothesis) and then look for a method to solve it. The successful projects start with
some new method and then look for a problem.
Let’s take Bohnenblust’s advice. Let’s start with linear programming and look for a
problem. Here’s a good one: the problem of moments in probability theory.
Suppose we are given a collection of real-valued continuous functions aistd for t [ Rp.
We are given a closed set V , Rp, and we’re given a collection of real numbers bi.
The problem is to find a probability distribution function x(t) satisfying the moment
equations
15
 EV
aistd dxstd 5 bi for all i

where we require dxstd ≥ 0 and

EV
dxstd 5 1.

This problem has many applications in geophysics and in other sciences. It has an
extensive mathematical theory (see, for instance, Shohat and Tamarkin [30]). So what is
left for you and me to do here? Well, for one thing, we could devise a good numerical
method. At least, that will please our colleagues in geophysics.
Suppose we’re given a finite number of moments, which is the usual case in
applications. And suppose we use some numerical scheme to approximate the integrals
by finite sums. Then we get a finite set of linear equations in a finite set of unknowns:
n

oa x 5 b
j51
ij j j si 5 1, . . ., md.

Now we’re looking for the numbers x1, . . ., xn; they will constitute a finite set of
probabilities, satisfying
n

o x 5 1,
j51
j xj ≥ 0.

So we want to solve m 1 1 linear equations in n unknowns xj ≥ 0. Ah! We recognize

a problem in linear programming. For this we have an existence theorem, the Farkas
theorem, and a numerical method, the simplex method.
The simplex method will tell us if no solution exists, or it will compute a solution x if
solutions do exist. For n > m 1 1 we can’t expect the solution x to be unique. We are
free to impose any minimum condition of the form
n

o c x 5 minimum.
j51
j j

We note that the original problem with a finite number of moments usually doesn’t have
a unique solution xstd, so the freedom to impose an extra condition is physically natural
and mathematically necessary.
Fine, you say. All right for some people but not for you. You are a pure mathematician,
and numerical methods bore you. What you’d like is a little solid theory—something
you can get your teeth into.
OK, I’m with you. Let’s prove a great theorem together. Let’s give a new, elementary
proof of a famous theorem of F. Hausdorff [14]. The proof will use a method of
mathematical economics, the Farkas theorem.
Hausdorff studied the moment problem

(3) E 0
1
tk dxstd 5 bk sk 5 0, 1, . . .d.

16
He asked this question: Which infinite sequences HbkJ are the moments of a probability
distribution xstd on the interval 0 ≤ t ≤ 1? He called those sequences moment
sequences.
Certainly b0 5 1, since we require e dxstd 5 1. Also we must have

E0
1
fstd dxstd ≥ 0

for all continuous functions fstd ≥ 0. Setting fstd 5 tjs1 2 tdk, we get the necessary
condition

Eo1 k

0 v50
s2 dv 1kv2t j1v dxstd ≥ 0,

which says this about the moments:

o s2 d 1v2b
k k
v
j1v ≥ 0 sj, k ≥ 0d.
v50

A sequence HbiJ with this property is called completely monotone. If we define the
difference operator D by Dbi 5 bi11 2 bi, the last formula says
s2 dkDkbj ≥ 0 sj, k ≥ 0d.
Hausdorff’s theorem says: If b0 5 1, the sequence b0, b1, b2, . . . is a moment sequence if
and only if it is completely monotone.
We’ve already proved the only if part. To prove the if part, let’s assume the sequence
HbiJ is completely monotone, with b0 5 1. Now we must find a p.d.f (probability
distribution function) xstd satisfying the moment equations (3).
Suppose we can solve the system of moment equations

(i) E 0
1
tk dxnstd 5 bk sk 5 0, . . ., nd

for each finite n. Then the p.d.f.’s xnstd have a subsequence that converges to a p.d.f. xstd
at all points of continuity of the limit xstd. Then xstd satisfies all the moment equations
(3), and we’re done.
So the required p.d.f. xstd exists unless some finite system (i) is unsolvable. But the
system (i) is a finite linear system for an unknown dxnstd ≥ 0. A simple extension of the
Farkas theorem says this: The system (i) is unsolvable for a p.d.f. xnstd if and only if
there exist numbers yo, . . ., yn satisfying
n

(ii) o yt
k50
k
k ≥ 0 s0 ≤ t ≤ 1d
n

o yb
k50
k k < 0.

We must show that this is impossible.

Suppose (ii) is true. Define the polynomial fstd 5 oyktk. Then Taylor’s theorem says
yk 5 f skds0dyk!
17
As a limit of difference quotients, this equals
yk 5 lim D«kfs0dys«kk!d,
«→0
where D« f std 5 fst 1 «d 2 fstd. Setting « 5 1yN, we deduce

` k2
1 N
yk 5 lim D f s0d.
k
«
N→

The second part of (ii) says oykbk < 0, and so for large N we must have

o 1 k2sD
n N k
« f s0dd ? bk < 0.
k50

The upper limit, n, may be replaced by a larger integer, N, since an nth degree
polynomial fstdsatisfies D«k f std 5 0 for k > n. Now we rearrange the last sum to obtain
the inequality

o f1N21 j2 ? s2 d
N j N N2jDN2jb < 0.
j
j50

But (ii) says f ≥ 0, and the completely monotone sequence HbiJ satisfies s2 dkDkbj ≥ 0,
so all terms in the last sum are nonnegative, and we have a contradiction. The Farkas
alternative (ii) is impossible.
Therefore, the alternative (i) is true: every finite system of moment equations (i) is
solvable. It follows that the infinite system (3) is solvable, and so we have proved
Hausdorff’s theorem.
This theorem is important in probability theory. As William Feller said, “Its discovery
has been justly celebrated as a deep and powerful result.” (See [9], p. 226.)
As you’ve just seen, the mathematical methods of economics have striking applications
to the rest of mathematics. As you might have feared, I could go on talking to you
forever. I could tell you about applications to ill-posed boundary-value problems of
partial differential equations. But I manfully refrain; you have already heard enough. By
now, I hope you will agree with me: these problems and methods of economics are
valuable, and they are fascinating.
References
1. Kenneth Arrow, Social Choice and Individual Values, 2nd ed., Yale University Press, New
Haven, 1963.
2. Robert G. Bland, The allocation of resources by linear programming, Sci. Amer., 244 (1981)
126–144.
3. Alfred Broaddus, Linear programming: a new approach to bank portfolio management,
Monthly Review, Federal Reserve Bank of Richmond, Virginia, vol. 58 (1972) 3–11.
4. Donald J. Brown, Nonstandard economics: a survey, to appear in Studies in Mathematical
Economics, S. Reiter, editor, Math. Assoc. of America.
5. George B. Dantzig, Constructive proof of the min-max theorem, Pacific J. Math., 6 (1956)
25–33.
6. _____, Linear Programming and Extensions, Princeton Univ. Press, 1963.
7. Gerard Debreu, Theory of Value, Yale University Press, New Haven, 1959.

18
 8. William Feller, An Introduction to Probability Theory and Its Applications, vol. 1, 3rd ed.,
Wiley, New York, 1968.
9. _____, An Introduction to Probability Theory and Its Applications, vol. 2, 2nd ed., Wiley,
New York, 1971.
10. Joel Franklin, Methods of Mathematical Economics, Springer-Verlag, New York, 1980.
11. John Kenneth Galbraith, Economics, Peace, and Laughter, essays edited by Andrea D.
Williams, Houghton Mifflin Company, New York, 1971.
12. O. A. Gross, Preferential arrangements, this MONTHLY, 69 (1962) 4 –8.
13. R. F. Harrod, The Life of John Maynard Keynes, Avon Books, New York, 1951.
14. F. Hausdorff, Momentprobleme für ein endliches Intervall, Math. Z., 16 (1923) 220–248.
15. S. Kakutani, A generalization of Brouwer’s fixed-point theorem, Duke Math. J., 8 (1941)
457–458.
16. L. V. Khachian, A polynomial algorithm in linear programming. Soviet Math. Dokl., 20
(1979) 191–194 (Amer. Math. Soc. translation).
17. Alan P. Kirman and Dieter Sondermann, Arrow’s theorem, many agents, and invisible
dictators, J. Econom. Theory, 5 (1972) 267–277.
18. Victor Klee and George Minty, How good is the simplex algorithm? Inequalities, 3 (1972)
159–175, Academic Press, New York.
19. H. W. Kuhn, The Hungarian method for the assignment problem, Naval Res. Logistics
Quart., 2 (1955) 83–97.
20. W. A. J. Luxemberg and K. Stoyan, The Theory of Infinitesimals, Academic Press,
New York, 1976.
21. G. S. Maddala, Econometrics, McGraw-Hill, New York, 1978.
22. John Milnor, Analytic proofs of the “hairy ball theorem” and the Brouwer fixed-point
theorem, this MONTHLY, 85 (1978) 521–524.
23. J. F. Nash, Equilibrium points in n-person games, Proc. Nat. Academy of Sciences, 36
(1950) 48–49.
24. John von Neumann, Zur Theorie der Gesellschaftsspiele, Math. Ann., 100 (1928) 295–320.
25. John von Neumann and Oskar Morgenstern, Theory of Games and Economic Behavior,
Princeton Univ. Press, 1944.
26. H. Nikaido, Convex Structures and Economic Theory, Academic Press, New York, 1968.
27. C. A. Rogers, A less strange version of Milnor’s proof of Brouwer’s fixed-point theorem,
this MONTHLY, 87 (1980) 525–527.
28. Paul Samuelson, Economics, 9th ed., McGraw-Hill, New York 1973.
29. Jacob Schwartz, Mathematics as a tool for economic understanding, pp. 269–295 in
Mathematics Today: Twelve Informal Essays, edited by L. A. Steen, Springer-Verlag,
New York, 1978.
30. J. A. Sholat and J. D. Tamarkin, The Problem of Moments, Mathematical Surveys, no. 1,
American Mathematical Society, New York, 1943.
31. G. F. Shove, The place of Marshall’s Principles in the development of economic theory,
Econom. J., 52 (1942) 294–329.
32. Eduard Stiefel, An Introduction to Numerical Mathematics, Academic Press, New York,
1963.
33. George Stiefel, The cost of subsistence, J. Farm. Econom., 27 (1945) 303–314.
34. J. J. Sylvester, A question in the geometry of situation, Quart. J. Pure and Appl. Math., 1
(1857) 79.
19
35. A. Takayama, Mathematical Economics, Holt, Rinehart, and Winston, 1974.
36. Philip Wolfe, The simplex method for quadratic programming, Econometrica, 27 (1959)
382–398.

20