The Introduction of Risk Into A Programming Model
The Introduction of Risk Into A Programming Model
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.
The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to Econometrica.
http://www.jstor.org
This content downloaded from 143.106.81.78 on Mon, 14 Sep 2015 15:12:15 UTC
All use subject to JSTOR Terms and Conditions
RUDOLF
J.
FREUND1
1. INTRODUCTION
CONSIDER
THE
v
0.
THE DESCRIPTION
OF A RISK
SITUATION
We will assume that the money outcome of a unit level of a process carried on
under risk conditions is in the form of a random variate which follows some
probability distribution. This distribution can be defined as representing some
measure of the degree of belief that particular outcomes will occur. This degree
of belief is, of course, a highly subjective concept, but it is conceivable that
1 This paper is based on a thesis presented to North Carolina State College, June, 1955.
The author wishes to acknowledge the help of Clifford Hildreth whose ideas form the background for a large part of this work.
253
This content downloaded from 143.106.81.78 on Mon, 14 Sep 2015 15:12:15 UTC
All use subject to JSTOR Terms and Conditions
254
RUDOLF
J.
FREUND
s,:N(,u,,
oa2)
with the covariance between the net revenue of two processes defined as ?j,
then the net revenue, r, will also be normally distributed:
r:N(,u'x,X'2;X),
where 2 is the matrix of variances and covariances of the si. Of course, it is
possible to assume any other distribution of the si, but due to the distribution
problems posed by the use of most other distributions and the lack of data either
to support or reject any assumed distribution the normal will be used in the
development of a risk program.
There does, however, exist one alternative to the assumption of a normal
distribution; this consists of the use of a discrete distribution. Thus, it is assumed
that a particular combination of unit level net revenues, say
(s
)I
(Si t)
S2
t)
. . .
s( t)
t =
l1t
2,
has a probability, pt, say. The probability distribution of net revenue due to a
program x can then be defined:
Pr(r Ix) = p((sPt))'x).
This formulation of the probability of net revenue will be used in an alternative
development of a risk program.
One feature of both methods of describing an uncertain outcome is that they
depend on the decision variable, the program x. Thus the entrepreneur making a
This content downloaded from 143.106.81.78 on Mon, 14 Sep 2015 15:12:15 UTC
All use subject to JSTOR Terms and Conditions
255
MODEL
A PROGRAMMING
decision about x must choose from among a set of probability distributions; the
number of distributions in the set is governed by the various possible values of x.
3.
THE EVALUATION
OF RISKY
ALTERNATIVES
The evaluation and consequent choice from among a set of probability distributions will here be accomplished by the use of the expected utility hypothesis
[3, 11]. According to this theory an entrepreneur behaves as though he places
some numerical utility, invariant through transformation of origin and unit of
scale, on each certain outcome and then acts to choose that probability distribution which exhibits the maximum (mathematical) expected utility.
In the case under discussion here, the outcomes are all in terms of net revenues.
Thus the utility of certain outcomes is equivalent to the utility of net revenue,
or the utility of money; the relationship between the outcomes and their utilities
becomes the utility of money function. In general this utility function describes
the nature of the entrepreneur, i.e., whether he is a gambler or whether he is
conservative, and how extreme is his inclination toward either of these characteristics [8].
One such utility function which we will find convenient to use is
y (r) - 1
where y is the utility and r is the net revenue. This function is convex everywhere
and indicates a conservative entrepreneur. The constant a indicates the entrepreneur's aversion to risk; the larger the value of a, the more conservative the
entrepreneur. If we now assume that net revenue is normally distributed, the
maximization of expected utility
E(u)
(1 -_ C)
e(r
)2/2a2
dr
a .
Remembering the form of the parameters of the distribution of the net revenue
in a programming model, the maximization of expected utility can be accomplished by maximizing
E(u*)3-s'x
2a ,x
normal integral multiplied by - exp [(a2o-2/2) - a,pl. Since the integral becomes a constant, the maximization of E(u) is equivalent to minimizing the exponent which is in turn
accomplished
by maximizing I
(a/2)cr2.
256
RUDOLF
J. FREUND
and
x > 0.
This is now a problem in quadratic programming [6, Ch. III] and [10], for which
a solution procedure has not yet been completely perfected. In the next section a
short computational example is presented and in the appendix there is an explanation of the procedure used and an attempt to show that this is an optimum
solution.
It was pointed out above that we can assume a nonparametric (discrete) distribution of the vector s. If we let s be a random variate and assume the utility
of money function
y(r) = r
br2
bx'E(ss')x,
where
E(s) = , Pr(s())s(t
and
E(s8') =EPr(S('I)P)S(ts'
The maximization of expected utility in the programming model is now also a
quadratic programming problem, viz.
to maximize E(u Ix)
subject to
and
Tx
x
<
dx'
bxFx'
v
O,
where d is the expression for E(s) and F is the quadratic form for E(ss').
The use of a quadratic utility function of money needs some clarification.
In the normal case this would not be an acceptable function as a quadratic has
a negative slope in some region where there exists a probability of r. In the discrete case, however, this region can be made to be outside the realm of possibility, i.e., in a region where the probability of r is zero.
4.
A COMPUTATIONAL
EXAMPLE
A PROGRAMMING
257
MODEL
TABLE I
THE MATRIXOF UNIT LEVELRESOURCE
REQUIREMENTS
Amounts needed by unit level of process
Scarce
resources___________
Scarce
resources
1
Early Irish
potatoes
Land, acres
January-June
July-December
Production Capital, $100
Period 1 (Jan.-April)
Period 2 (May-Aug.)
Period 3 (Sept.-Dec.)
Managerial labor, hours
Period 1
Period 2
Period 3
Unit level profit $100
4
Fall Cabbage
Availability
resources of
2
Corn
3
Beef
1.199
.000
1.382
1.382
2.776
2.776
.000
.482
60.
60.
1.064
-2.064
-2.064
.484
.020
-1.504
.038
.107
-1.145
.000
.229
-1.229
24.
12.
0.
5.276
2.158
.000
4.836
4.561
4.146
.000
.000
.000
.000
4.198
13.606
799.
867.
783.
1.000
1.000
1.000
1.000
TABLE II
THE OPTIMuM PROGRAM
Process
1.
2.
3.
4.
Potatoes .22.14
Corn..00
Beef (pasture) .11.62
Fall cabbage.57.55
Unit Levels
Acres
26.55
.00
32.26
27.74
258
RUDOLF
J. FREUND
Ideally the variances (and covariances) of the unit level net revenues should
measure the variabilities in the net revenue of these processes under conditions
of constant price levels and constant managerial practices. Since such data are
not only difficult to obtain, but are also unavailable for the type of farm which
is being analysed, approximate procedures have been used to obtain estimates
of the elements of the matrix M.First, it was assumed that only prices and yields
of the outputs are random variates; prices and quantities of all inputs, such as
fertilizer, tractor costs, etc., are assumed fixed at given levels. This assumption
is not unreasonable for the type of farm we are considering here, but would not be
realistic for a corn-hog farm, for example, where the input, corn, certainly will be
subject to price fluctuations. Secondly, prices and yields used to obtain estimates
of past unit level net revenues consisted of adjusted state averages. These past
unit level revenue estimates were used to obtain the variance-covariance matrix 2.
The main disadvantage of the use of state averages is that it definitely underestimates the variance since the basic data are already averages. This is especially
serious in the case of the yields where, especially due to the irregular pattern of
summer storms, the individual farm variabilities are averaged out over larger
areas. Some limited data were examined which indicated that the ratio of farm
variances to state average variances of yields may be as much as three or four
to one.
The estimation of the risk aversion constant a is a purely subjective task,
and any chosen value is exceedingly difficult to defend. It would seem to this
author that for a farm of the size of the one in this article, a value somewhere
between 1/2500 and 1/5000 is reasonable. However, to adjust for the fact that
the variances were underestimated, a value of 1/1250 was used for a.
The unit level variance covariance matrix, 2, is presented in Table III. The
quadratic programming problem:
to maximize s'x
2 x'Ex
subject to Tx < v
and
x > 0
was solved by the use of the Kuhn and Tucker [10] minimax theorem and a
solution procedure developed by Hildreth [9]. The procedure is more explicitly
presented in the Appendix.
TABLE III
UNIT LEVEL VARIANCE-COVARIANCE MATRIX
Crop
Potatoes ............
Corn
Beef
Fall Cabbage ........
Potatoes
Corn
Beef
Fall Cabbage
7304.69
903.89
620.16
-688.73
-471.14
1124.64
-1862.05
110.43
750.69
3689.53
This content downloaded from 143.106.81.78 on Mon, 14 Sep 2015 15:12:15 UTC
All use subject to JSTOR Terms and Conditions
259
MODEL
A PROGRAMMING
TABLE IV
COMPARING THIE RISK AND THE NO-RISK
Item
OPTIMUM PROGRAMS
Risk Program
No-Risk Program
10.31
26.75
2.68
32.35
22.14
.00
11.62
57.54
12.36
36.97
7.44
15.59
7209
5369
1517
26.55
.00
32.36
27.73
9131
4393
2434
Process Intensities
Unit levels
Potatoes
Corn
Beef
Fall cabbage
Acres
Potatoes
Corn
Beef
Fall cabbage
Expected net revenue, dollars
Expected utility, arbitrary units
Standard deviation of net revenue, dollars
The resulting optimum risk program is presented in Table IV. The no-risk
program is given alongside for comparison. The results are about what were to
be expected. The contribution to the total net revenue of the high risk crops of
potatoes and cabbage was reduced from 87.3% to 59.2%, in addition to which
more "safety" is achieved by the fact that the total net revenue does not depend
largely on a single crop. It is true that the net revenue has been decreased, but
so has the standard deviation of net revenue. It may be argued that the reduction
in the standard deviation was bought at a rather large expense in net revenue;
this is due to the fact that the risk aversion constant a was given a rather high
value to counter the expected underestimation of the variances.
The quadratic programming procedure used produces a set of imputed prices
to the limited resources similar to those produced in the linear programming
procedure ([6], Ch. H, section 6, and Ch. III, section 4). These imputed prices
TABLE V
IMPUTED PRICES OF THE RISK AND NO-RISK
Risk Program
PROGRAMS
No-Risk Program
Item
Imputed Prices in Dollars
Land, acres
Spring
Fall
Production Capital $100
Period 1
Period 2
Managerial Labor, hours
Period 1
Period 2
Period 3
.00
32.93
.00
34.74
65.96
.00
94.01
.00
.00
.00
.00
.00
.00
6.10
This content downloaded from 143.106.81.78 on Mon, 14 Sep 2015 15:12:15 UTC
All use subject to JSTOR Terms and Conditions
260
RUDOLF
J. FREUND
differ from those in the linear case in that the total imputed cost, sum of the
products of the imputed prices and the quantities used is less than the total net
revenue. The imputed prices do, however, indicate by what amount the function
to be maximized can be increased by increasing the supply of a scarce resource.
In the risk programming model, an increase in the function to be maximized will
be in dollar terms. This occurs since an increase in the function
y =
2
&-
ao
261
MODEL
A PROGRAMMING
OF A QUADRATIC
PROGRAMMING
PROBLEM
Computational methods for quadratic programming have not yet been completely
developed. One method is described below and it is shown that the resulting solution is
indeed optimum. One other method, suggested by Charnes and Lemke [5], approximates
curved lines of separable quadratic or other convex nonlinear functions by series of connected straight line segments, whence the problem is solved by linear programming techniques.
Kuhn and Tucker [10] proved that the quadratic programming problem
to maximize s'x
x'Ex
subject to Tx < v
is equivalent to the minimax problem
maximize with respect to xlx)
minimize with respect to uf
subject to u > 0,
where u is a vector containing as many elements as T has rows.
This is not identical with the quadratic programming for the risk problem which is to
maximizes'x
subject to Tx < v
and x > 0.
Some redefinitions will, however make the risk programming problem amenable to the
minimax approach. First let %aS becomeE* and then define
T*=(
I)'
and v*=(0
where -I is negative identity matrix and 0 is a vector of zeros; T* and v* have the same
number of rows. Then, the risk programming problem can be stated
maximize s'x -x'*x
subject to T*x 6 v*,
262
RUDOLF
J. FREUND
ever, only proved the procedure for cases where the matrix T (or T*) is of full row rank.
In the risk programming problem, however, the matrix T* cannot be of full row rank due
to the fact that the negative identity matrix forces T* to have more rows than columns.
The procedure is quite adequately presented in Hildreth's paper, whence the presentation here will be quite brief and will serve only to facilitate the exposition of the method
used to check the optimality of the solution obtained. The minimax problem is
maximize with respect to x
ux
4+(x, u)
s'x -x'Z*'
+ 4'(v -T*x)
subject to u > 0.
Since the maximization with respect to x is unrestricted, this can be accomplished by taking
the derivative, setting it equal to zero and substituting back. The problem then remaining
is
minimize with respect to u: +*(u)
b'u + u'Cu
subject to u > 0,
where
kc =
b' = V'-
he*ts
*-1 T*'
and
C
= T*E*-l
T*'.
A PROGRAMMING
MODEL
263
This procedure was used to check the optimality of the risk program in the text. In the
solution, only two restrictions were found to be effective: Fall land and period 1 capital.
Since there are four active processes, the T* matrix will be of full row rank with up to four
restrictions. Thus two more restrictions, those of spring land and nonnegativity of the
beef process were added, more or less as a check. The procedure used followed directly the
outline by Hildreth [9], and the results were identical with those obtained by the original,
nine-restriction problem, whence we can assume that the solution is optimal.
REFERENCES
[1] ARROW,KENNETH J.: "Alternative Approaches to the Theory of Choice in Risktaking Situations," Econometrica, 19 (1951), 404-437.
[21 BABBAR, M. M.: "Distributions of Solutions of a Set of Linear Equations," Journal
of the American Statistical Association, 5 (1955), 854-869.
[31 BERNOULLI, D.: "Specimen theoriae novae de mensura sortis," Commentariiacademiae
scientiarium imperiales Petropolitanae, 5, 175-192. Translated by Dr. Louise Sommer, Econometrica,22 (1954), 23-36.
[4] CHARNES, A., W. W. COOPER, AND A. HENDERSON: An Introductionto Linear Programming. New York: John Wiley and Sons, Inc., 1953.
[5] CHARNES, A., AND C. E. LEMKE: Computational Theory of Linear Programming, II.
Minimization of Non-linear Convex Functionals. Graduate School of Industrial
Administration, Carnegie Institute of Technology, Pittsburgh, Pennsylvania,
1954.
[61 DORFMAN, ROBERT: Application of Linear Programming to the Theory of the Firm.
Berkeley: University of California Press, 1951.
[7] FREUND, R. J.: "The Introduction of Risk into a Linear Programming Model," unpublished thesis, North Carolina State College, 1955.
[8] FRIEDMAN, MILTON, AND L. J. SAVAGE: "The Utility Analysis of Choices Involving
Risk," The Journal of Political Economy, 56 (1948), 279-304.
[9] HILDRETH, C. G.: "Point Estimates of Ordinates of Concave Functions," Journal
of the American Statistical Association, 49 (1954), 598-619.
[101KUHN, H. W., AND A. W. TUCKER: Nonlinear Programming. Second Berkeley Symposium on Mathematical Statistics and Probability. Berkeley: University of California Press, 1950.
[11] VON NEUMANN, J., AND 0. MORGENSTERN: The Theory of Games and Economic Behavior, second edition. Princeton: Princeton University Press, 1947.
[12] SAVAGE, L. J.: The Foundations of Statistics. New York: John Wiley and Sons, Inc.,
1954.
This content downloaded from 143.106.81.78 on Mon, 14 Sep 2015 15:12:15 UTC
All use subject to JSTOR Terms and Conditions