Professional Documents
Culture Documents
Managing Editors:
J.P. Ancot, Netherlands Economic Institute, Rotterdam, The Netherlands
A.J. Hughes Hallett, University of Newcastle, U.K.
Editorial Board:
F.G. Adams, University of Pennsylvania, Philadelphia, U.S.A.
P. Balestra, University of Geneva, Switzerland
M.G. Dagenais, University of Montreal, Canada
D. Kendrick, University of Texas, Austin, U.S.A.
J.H.P. Paelinck, Netherlands Economic Institute, Rotterdam, The Netherlands
R.S. Pindyck, Sloane School of Management, M.I.T., U.S.A.
H. Theil, University of Florida, Gainsville, U.S.A.
W. Welfe, University of Lodz, Poland
For a complete list of volumes in this series see final page of this volume.
Developments of Control Theory
for Economic Analysis
edited by
1987
KLUWER ACADEMIC PUBLISHERS .1.
DORDRECHT I BOSTON I LANCASTER ,~
IV
Distributors
jor the United States and Canada: Kluwer Academic Publishers, P.O. Box 358.
Accord Station, Hingham, MA 02018-0358, USA
jor the UK and Ireland: Kluwer Academic Publishers, MTP Press Limited.
Falcon House, Queen Square, Lancaster LAI lRN, UK
jor all other countries: Kluwer Academic Publishers Group, Distribution Center.
P.O. Box 322, 3300 AH Dordrecht, The Netherlands
Copyright
PREFACE
Giovanni Castellani
Rector of the University of Venice
TABLE OF CONTENTS
Preface V
G. Castellani
In trod uction XI
C. Carraro and D. Sartore
INTRODUCTION
C. Carrara and D. Sartore (eds). Developments of Control Theory for Economic AnalysIS
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht·
XII
and point out the necessity of new achievements which will be more
extensively taken up in the second section of the book.
References
CHAPTER 1
Gregory C. Chow
Princeton University, USA
Jan Tinbergen and Ragnar Frisch received the first Nobel Prize
in economic science in 1969 for their contributions to the modelling
of dynamic economic relationships. Macroeconometric models
form the basis for the study of macroeconomic policy and for the
formulation of good policies. Tinbergen (1952, 1956) provided a
conceptual framework for quantitative policy analysis. He dis-
tinguished between economic policies which constitute basic
reforms of economic institutions and those which attempt to
achieve specified objectives given the economic institutions. Quan-
titative analysis was applied to determine the latter kind of policies.
An important ingredient in the analysis is an econometric model
which is used to evaluate the effects of alternative economic policies.
C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic AnalysIS
© 1987 Martinus NiJhoff Publishers (Kluwer), Dordrecht-
4
(4)
(5)
where
-(C'HTC)-IC'HTA (6)
Equation (5) gives the optimal feedback rule x T for the last period.
9
Because the second line of (9) is identical with the last expression of
(4) with Treplaced by T - 1, the solution for XT_I is identical with
(5) with T replaced by T - 1, where GT_ I and gT-I are defined by
(6) and (7) respectively with a similar change in time subscripts.
Accordingly, VT _ I will be given by (8) with the subscripts T replaced
by T - 1.
When we attempt to solve the problem for the last three periods,
we observe that xT and XT _ I have been found that would yield the
minimum expected loss VT -I for the last two periods and that, by
the principle of optimality, we need only minimize
VT- 2 = E T- 3 [(YT-2 - aT- 2)'KT- 2(YT-2 - a T- 2) + VT_ I ]
with respect to X T - 2 , and so forth. At the end of this process, we find
XI = Glyo + gl as the optimal policy for the first period, and the
associated minimum expected loss VI for all periods (or from
10
(14)
where Yfl is a random vector with zero mean and covarance matrix
Q and is statistically independent of the vector VI in (2). By the
Kalman filter of Kalman (1960), one can estimate the mean and
covariance matrix of y, recursively from the mean and covariance
matrix of YI_I and the new observation SI' Given the mean and
covariance matrix of Yo at time 0, successive observations SI' S2, . . .
are used to estimate the means and covariance matrices of YI,
12, .... The optimal policy is given by the same feedback rule as
before, except that in the feedback control equation for XI' YI_I is
replaced by £I-IYI-I as estimated by the Kalman filter. This result
is known as the separation theorem since the estimation of £t-IYI-I
can be performed separately from the derivation of the optimal
feedback rule. It is also related to the optimality of the certainty-
equivalent method since the optimal feedback rule is the same as the
rule derived when the state vector YI is known for certain.
Kalman filtering has found many applications in econometrics
and in macroeconomic policy analysis. One of the most important
is the estimation of parameters which are time-varying. The
simplest example is a regression model
(15)
PI'
In macroeconomic applications of optimal control methods, one
finds that frequently the simultaneous-equations models used
are nonlinear. Let us write the ith structural equation in such a
12
model as
(17)
where, as in equation (1), Yt and Xt denote vectors of endogenous
and control variables respectively; 8 11 is a random disturbance and
)'11 is a vector consisting of exogenous variables not subject to
control and parameters. A system of nonlinear structural equations
can be written as
(18)
with <I> denoting a vector function. Analytical solution to the prob-
lem of minimizing the expectation of (3) subject to the nonlinear
model (19) is difficult to obtain. In econometric practice, two
approximate solutions have been proposed, one taking a feedback
form and the other an open-loop form.
To obtain an approximately optimal feedback rule, one can work
with a certainty equivalent model by omitting the random term 8 t
in (18), linearize the resulting model, and find the optimal feedback
rule for the linearized model. Specifically, after dropping 8 t from
(18) and given the initial state Yo and the vector At for t,
t = 1, ... , T, one assumes a tentative policy path x~ and solves
the certainty-equivalent model
Yt = <I>(YI'Yt_"xI'At) (19)
for Yt. The solution l satisfies
y~ = <I>(y~, Y~_I' x~, )'t) (20)
A popular method to solve the nonlinear simultaneous equations
for YI is the Gauss-Siedel method. One then linearizes the certainty-
equivalent model about the path (y~, x?) to yield
YI - y~ = <l>lt( YI - y~) + <l>2t( Y{-I - Y~-I) + <l>3{(X - x?) + 81
(21)
to which we have added the random disturbance 8{. Solving the
linearized structural equations (21) one obtains a linearized reduced
form
(22)
As before, optimal feedback rules to minimize the expectation of (3)
13
subject to (22) can be derived, which are the same as (5), (6), (7),
(10), (11) and (12) with suitable time subscripts added to A and C.
Once the optimal rule for XI is derived, a new tentative path for x~
and y~ can be calculated for the certainty-equivalent model. A
second optimal rule for XI can be derived, and the iterative process
goes on until convergence. The resulting feedback rule is optimal
for the nonlinear certainty-equivalent model (19), as it can be easily
shown (Chow, 1975, pp. 289-295). It is presumably nearly optimal
for the original model (18).
A second approach is open-loop. Using the nonlinear certainty-
equivalent model (19), a solution y~ corresponding to any open-
loop policy x~ (t = 1, ... , T) can be calculated by the Gauss-
Siedel or another numerical method. The loss associated with
(l, x~) can be evaluated using the loss function. Thus loss is a
function of the open-loop policy. This function can be minimized
with respect to XI' • . . , X T by a gradient method. This resulting
open-loop solution is optimal for the certainty-equivalent model. It
is identical to the solution obtained by the first approach in the
certainty case since both are optimal (provided that the solution is
unique). Computationally, the second approach is efficient if the
number of control variables and the number of periods T in the
planning horizon are small, the total number of variables in the
minimization problem being the product of the two. As the number
of periods T is increased, the number of variables increases propor-
tionally by the second approach and the computational time in-
creases more than proportionally, whereas, by the first approach,
the computational time increases only proportionally. Having
available the feedback control equations obtained by the first
approach, one can conveniently study the dynamic properties of the
model under control.
If uncertainty is introduced by retaining the random disturbance
ef' both the closed-loop feedback approach and the open-loop
approach have to be modified. The former is discussed in Chow
(1976) and the latter is discussed in Fair (1974). In the latter case, one
has to minimize the expectation of the loss function with respect to
XI' . . . , X T where the expectation may be evaluated by stochastic
simulations of the nonlinear stochastic model (18). If uncertainty is
further introduced by allowing for the errors in estimating the
parameters of the model (18), we have a dual control problem. The
term dual control represents two aspects of setting the optimal
14
control policy, one to steer the state vector Yt to its target and the
second to set x t to improve the future observations of X t and Yt for
the purpose of obtaining better estimates of the parameter, thus
improving future control of Yt. Much work on dual control has
been done in the control literature. In the economics literature, the
works include MacRae (1972), Tse (1974), Chow (1975, Chapter
11), Kendrick (1981) and Norman (1984), among others
Given this policy rule, the environment facing the economic agents
becomes
YI (A + C 2G2)YI_1 + Clx lt + bl + C2g 21 + VI
References
Kendrick, David (1981): Stochastic Control for Economic Models. New York:
McGraw-Hili Book Company.
Lucas, Robert E., Jr. (1976): "Econometric Policy Evaluation: A Critique," in
K. Brunner and A. H. Meltzer (eds.), The Phillips Curve and Labor Markets.
Carnegie-Rochester Conference Series on Public Policy, Vol. 1. Amsterdam:
North-Holland Publishing Company.
MacRae, R. C. (1972): "Linear Decision with Experimentation." Annals of
Economic and Social Measurement, 1,437-448.
Marschak, J. (1953): "Economic Measurements for Policy and Prediction," in
W. C. Hood and T. C. Koopmans (eds.). Studies in Econometric Method,
Cowles Commission Monograph 14. New York: John Wiley & Sons, Inc.
Miller, M. and M. Salmon (1984): "Dynamic Games and Time Consistent Policy
in Open Economics." Mimeo.
Muth, J. F. (1961): "Rational Expectations and the Theory of Price Movements,"
Econometrica, 29, 315~335.
Norman, A. L. (1984): "Alternative Algorithms for the MacRae OLCV Strategy,"
Journal of Economic Dynamics and Control 7, 21~38.
Phillips, A. W. (1958): "The Relation Between Unemployment and the Rate of
Change of Money Wage Rates in the United Kingdom, 1861~1957," Economica,
25, 283~299.
Pindyck, Robert S. (1973): Optimal Planning for Economic Stabilization. Amster-
dam: North-Holland Publishing Company.
Sargent, T. J. (1979): Macroeconomic Theory. New York: Academic Press.
Simon, H. A. (1956): "Dynamic Programming Under Uncertainty with a Quad-
ratic Criterion Function," Econometrica, 24 (I), 74-81.
Theil, H. (1958): Economic Forecasts and Policy. Amsterdam: North-Holland
Publishing Company.
Tinbergen, Jan (1952): On the Theory of Economic Policy. Amsterdam: North-
Holland Publishing Company.
Tinbergen, Jan (1956): Economic Policy: Principles and Design. Amsterdam:
North-Holland Publishing Company.
Tinbergen, Jan (1969): "The Use of Models: Experience and Prospects," Noble
Prize Lecture delivered in Stockholm, December 1969; American Economic
Review, 71 (6), 17~22.
Tse, Edison (1974): "Adaptive Dual Control Methods," Annals of Economic and
Social Measurement, 3, 65-68.
Wan, Henry Y., Jr. (1983): "The New Classical Economics ~ A Game-Theoretic
Critique." Mimeo. Forthcoming in George Feiwel (ed.), Issues in Macroecon-
omics and Distribution. New York: The Macmillan Co.
Whittle, P. (1963): Prediction and Regulation by Linear Least Square Methods.
New York: D. Van Nostrand Co.
21
CHAPTER 2
R. Conti
University of Florence, Italy
o
Many models in Economics, as well as in other disciplines, can be
represented by a family of differential equations
dx/dt = Ax + c (A, c)
depending on a parameter c, called control.
The choice of the control within a given set determines the
evolution of the state x in time t, once the dynamics, i.e. the operator
A, and an initial state XO are given. Therefore (A, c) is called a
control process.
For a given XO the set of points x(t, xO, c) corresponding to all the
t > 0 and all the possible choices of c is a reachable set, or, more
precisely, the set of points reachable along the trajectories of (A, c)
issuing from XO at time t = O.
The study of some properties of the reachable sets, generally
known as controllability, is the object of the present talk.
C. Carrara and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-
22
where
I, identity,
as usual.
From (1.1) it is clear that without some restrictions on the choice
of c every reachable set would be the whole of IRn. So we shall
assume that every value c(t) has to belong to some proper subset f
of IRn . On the other hand it is also clear that f cannot be empty and,
in fact, it must consist of two points at least.
From now on we shall speak of the control process (A, c) in terms
of the pair (A, r).
holds.
When
W(A, r) = [Rn
holds then every point can be reached from O. Note that this does
not mean that every point can be reached from any other point like
in the case of complete controllability.
Property (R)o can be labelled as global reachability from O.
Changing A into - A we have
r is bounded,
but (3.1) is no longer necessary when r is unbounded, as examples
show.
Therefore it remains an open question that of determining some
condition, let us call it (x), less restrictive than (3.1), such that
(C)O,loc plus (x) ~ (R)o,
independently of (HI).
independently of (H2)'
hold. It may happen, in particular, that (5.1) hold for all t > O.
This means that for each t > 0 there is an open set Qo,{ contain-
ing 0 such that we can go from any point XO E Qo,{ to any other point
Xl E 0 0,/ along a trajectory of (A, c) in time ~ 2t.
Some authors call O-autoaccessibility this special kind of local
controllability. A characterization of pairs (A, r) having this
property is the following, due to R. M. Bianchini [11]. Let con r
denote the conic hull of r. Then (A, r) is O-autoaccessible if and
only if the pair (A, con r) is instantly controllable, i.e., if and only
if (2.2) holds with r replaced by con r.
27
Obviously
W(A, r) c W(A, r), VA, r
so that (R)o implies
W(A, r) = [Rn.
29
Simple examples show that (R)o does not imply (R)o' Take for
References
CHAPTER 3
1. Introduction
C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-
32
assures only that the system will "hit" a given target position in a
given moment of time, but it does not say anything about the
behaviour of the system after that moment. In economics it is not
sufficient to reach a desired position, it is also important to be able
to stay there. It would in fact be useless, for example, to bring
employment to a given desired level at a given moment ifit were not
possible to keep that level subsequently. Path controllability assures
instead that the system will remain in the position reached for a
given period of time (or that it will follow, in that period, a pre-
assigned trajectory). This is the reason why path controllability is
often considered as the only controllability concept that is relevant
for economic applications and as the dynamic generalization of
Tinbergen's theory.
It is nevertheless possible to show, as we shall do in the next
paragraph, that path controllability is only one of the ways of
generalizing Tinbergen's theory in a dynamic context, and that
point controllability can be also considered a dynamic extension of
the same theory. It is just the dynamic framework in which the
economic policy problem is analysed which allows to generalize
Tinbergen's theory in two different but equally important ways.
(1)
As we have just seen, the static equilibrium solution of this
system is given by equation (5), that is by
(5)
and, therefore, Tinbergen's static system (2) can be considered as
the equilibrium position of the dynamic system (1), such that
A = Ap and B = Bp. It is possible in this way to specify a direct
37
x(l) = Ax + Bu
(7)
y = Cx
5. Path Controllability
Consider the dynamic system in state space form (7). Taking the
Laplace transforms of both members we have
sX(s) - x(O) = AX(s) + BU(s) (10)
But to say that the system described by eq. (1) is right invertible
is the same as saying that the dual of system (1) is left invertible
since the transfer matrix of the dual system is defined as (See [9] and
[18])
(15)
where the prime denotes transposition, that is
G'(s) = B'(sI - A')-' C' (16)
CB 0 .............. 0
CAB CB . . . . . . . . . . . . . O
N (19)
CB
CA 2n - 1 B
We notice that being G(s) an m x r matrix if follows, from
condition (17), that:
40
o o. CB.
Let us now assume that Tinbergen's condition (3) is not satisfied
(i.e., rank [Bp] < m). It can be shown that the path controllability
condition (21) and, consequently, condition (22) can be satisfied
even if condition (3) is not. It is in fact possible to have rank
[C(s! - A)-I B) = m even if rank [Bp] < m; since rank [C) = m,
rank [(sf - A)-I] = n (n ~ m), and since we can have rank
[B) = m also when rank [Bp] < m, it is possible to have rank
[GD(s)] = m. 8
The same conclusion holds as far as condition (22) is concerned.
It is in fact possible to have rank [ND ] = (n + l)m even when rank
[Bp] < m; as we have just seen, we can have rank [B) = m even
when rank [Bp] < m. So we can also have rank [CB] = m even
when rank [Bp] < m. Given the form of the matrix ND this would
be sufficient to have rank [ND ] = (n + l)m.
Therefore,Tinbergen's condition (3) is not a necessary condition
for path controllability of continuous time dynamic systems.
41
7. Concluding Remarks
Notes
I. This paper has been prepared in the context of a research financed by the
Consiglio Nazionale delle Ricerche. I wish to thank G. Gandolfo and A.
Isidori for helpful comments.
43
y Cx + Du
where D E RW. This assumption does not alter the results obtained in the text,
but it eventual1y strengthens them. (See note (8).)
7. This assumption could be dropped if we consider the target vector yes) as the
difference between the same vector Yes) and the Laplance transform of the free
response of the dynamic system (I) (i.e. the trajectory which is the solution of
the corresponding homogeneous system). This is the approach followed in [4].
°
8. Note that in the case in which the term BouP(t) =1= in system (I), the transfer
function matrix of that system is
G(s) = C(sl - A) - 1 B + D
Therefore the result just stated in the text is strengthened by the presence of
the matrix D. In that case, in fact, it IS possible to have rank [G(s)] = m even
If rank [C(sl - A) -I B] < m.
References
26. Wohltmann, H. W.: "A note on Aoki's Conditions for Path Controllability of
Continuous-Time Dynamic Economic Systems". Review of Economic Studies
51 (1984), 343-349.
27. Wohltmann, H. W.: "Target Path Controllability of Linear Time-Varying
Dynamical Systems" IEEE Transactions on Automatic Control 30 (1985),
84-87.
28. Wohltmann, H. W., and W. Kromer: "A Note of Ruiter's Sufficient Con-
dition for Perfect Output Controllability of a Rational Expectations Model":
Journal of Economic Dynamics and Control 6 (1983),201-205.
29. Wohltmann, H. W., and W. Kromer: "Sufficient Conditions for Dynamic
Path Controllability of Economic Systems", Journal of Economic Dynamics
and Control 7 (1984), 315-330.
47
CHAPTER 4
David Kendrick
University of Texas at Austin, USA
C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NiJhoff Publishers (Kluwer), Dordrecht-
48
Response
Method Input Output
Time
EJ
fixed format
hours numbers
and numbers
1 1
.75 .9 .03 input
READ (5. 8) N. t1
FORt1AT(2I4)
8 program
READ(5. 10) BETA. ttJ. PI
10 FCRtIAT(3f5.2)
derivative of
F(l) = BETA * A(T)*K(T) **(BETA-l) function
Micro- B = .15 ;
1.1 = .90 ;
Computers I = .03 ;
III!Iximize
T -t (1-1.1)
(1) J = L [(1+.) --- c
t-1
t=l (1-1.1)
subject to
B
(2) a k
t t
(1962). The problem was solved fifteen years ago on an IBM 7094
with a Fortran code base on the conjugate gradient algorithm of
Lasdon, Mitter and Warren (1967). Part of an input file and some
code of this type is shown at the top of Figure 2. The data were in
fixed format and were read in with Fortran statements like those
shown in the figure. The equations of the model were represented
by functions for the derivatives like the one shown for function F(1)
which is the derivative of a Cobb-Douglas production function.
The derivatives were easy enough to do for this small model
50
maximize
T -t
(1) J = L [(1+0) ---
MATH t=1 (1-~)
subject to
B
(2) Yt a k
t t
t NEON
IMximize
.DE
.EQ
(1)
~~~~~~~~~~mark J ~ ;
subject to
.EQ
(2)
-----~~~Y sub t -;- a sub t k sub t sup beta;
.EN
+ MEL
L
(=1
2. Quadratic-Linear Models
Maximize
J = (x - x ) 'N (x - x )
N N N N
N-l
+ L [ (x - x ) Wk (x - x ) + (u - ~ ) • 1\ (u - ~ )
k=O k k k k k k k k k
subject to
x
k+l
..here
x = state vector (n x 1)
u = control vector (m x 1)
Figure 4. Mathematics for quadratic linear problem (QLP).
print controls
1 1
1. 014 .002
.093 .153 A
-.004
-. 100 B
-1.312
.448 c
1. O.
O. 1.
100. O.
1I
O. 100. N
1.
460. 1
113.1
463.551 461.021 410.530414.059 411.615 481.191484.806
113.948 114.803 115.664 116.531 111.405 118.286 119.113 )(
type of interface has been available for many years on the TROLL
system of the Centre for Computational Research in Economics
and Management Science (1978) and is coming in MODULECO by
Neopmiastchy and Rechenmann (1983). Moreover, it is a standard
kind of interface for microcomputer software.
One possible input form for the software of the future is shown
in Figure 8. The problem here is a dynamic control problem for the
college education financing plan of a family. There are two state
variables, namely the amount of money loaned to the family and
the amount of money in their bank account. Also, there are four
control variables, namely borrowing, savings, payments on the loan
and withdrawals for college expenditures.
55
Welcome to QLP.
element 1.2
.002
element 2. 1
.093
element 2. 2
.753
Input the B matr ix.
element 1. 1
-.004
Welcome to QlP.
a
1.014 0.002
0.093 0.753
b
-0.004
- O. 100
bank bank
account account
withdrawals
for
college
The software for this system records the location of each rec-
tangle and circle. This information potentially can be used to create
state equations which can then be written into the GAMS language.
It is the aspect of Lisa Draw and Mac Draw of keeping up with the
location of each object which is important. In contrast the popular
MacPaint software for the Macintosh does not keep up with each
object in the drawing and thus could not be used for this kind of
software.
Software which uses graphic input for optimal control problems
is thus a realistic possibility for the not too distant future. Similarly,
the output will not be the tables of numbers which we now use, but
rather graphs, charts and figures.
v.
00
I"E
0, ·1 loan (1+r) . / ' ..... loa"
c:J .
,~ ~ ~
r borrow!n, )'' ;
~
?t01 (
\ \\
~~.,;~~~ ') "\
'"
.~,
......... '-, ~
bank· ~~
acco::l · " . . . .r:,ank
account
References
CHAPTER 5
Josef Gruber
University of Hagen, FRG
*Gratefully acknowledged are many helpful discussions with Elke Petersen and
Michael Olbrisch who are now working on the research projects briefly described
in Sections 3 and 4 of this paper.
C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-Printed in the Netherlands
64
Group 1 0 0 0 0 I 0 0 0 0
Group 2 0 I 0 0 0 0 I 0 0 0
Group 3 0 0 I 0 0 0 0 I 0 0
Group 4 0 0 0 I 0 0 0 0 I 0
Group 5 0 0 0 0 0 0 0 0 I
Group 6 I I I 0 0 0 0 0
Group 6' 0 0 0 0 0 0 0 0
Group 7 0 I 0 0 0 0 0 0 0
Group 8 0 0 I 0 0 0 0 0 0
Group 9 0 0 0 I 0 0 0 0 0
( au aA
au au au au)'.
appr a~' I ' aPI' aG I ' aR I
This vector can be seen in row I of Table 3. Since only the direction
73
Table 3. One version of an answer matrix D' for Example I of the inquiry pattern.
Group Y; Al PI GI RI l2 A2 P2 G2 R2
Group 1 2 -1 -0.5 -0.1 0 0 0 0 0
Group 2 0 0 0 0 0 I 2 -1 -0.5 -0.1
Group 3 I 0 0 0 0 1 0 0 0 0
Group 4 0 I 0 -0.5 0 0 1.5 0 -I 0
Group 5 2 2 0 -1 0 2 2 0 -I 0
latter happens if
- the inquiry pattern consists only of a small number of groups of
objectives,
- the number of active objectives is small in comparison with the
number of objectives specified (i.e., if the inquiry pattern contains
many zero-columns).
- the answer matrix DS (which corresponds to a "well-specified"
inquiry pattern) contains extremely contradictory information
(without being inconsistent).
The results of numerous applications of the interactive computer
program combined with the Beckmann-Uebe-model can be sum-
marized as follows:
1. Rosinger's interactive algorithm for multiobjective optimization
is a suitable tool for macroeconomic programming - from the
theoretical point of view as well as in realistic decision situations.
2. In realistic decision situations the speed and safety of con-
vergence proved to be satisfactory: In most cases 10 to 12
iterations were sufficient to reach an/the optimal solution ("opti-
mal" in terms of the decision maker's preferences).
3. Problems arise in case of "extreme" situations (in the above
sense): The standard optimizing routine used for the solution of
the quadratic minimization problem that yields the vector a! in
(2.2) often did not reach a minimal value in the given (and
unalterable) number of iterations. It is planned to replace the
quasi-Newton-procedure by a procedure with a higher speed of
convergence.
79
5. Concluding Remarks
References
CHAPTER 6
Andries S. Brandsma
Erasmus University, Rotterdam, The Netherlands
Introduction
C. Carraro and D. Sarlore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NijhoJf Publishers (Kluwer), Dordrecht·
84
(16)
90
again towards the end of the planning period. This feature of the
risk sensitive decisions derived from the economic theory of risk
bearing aims to balance the portfolio of consumption and invest-
ment shares of national income. Its intention is supported by the
relative tightening of monetary conditions and the abandonment of
direct investment stimulation by the government. The results con-
firm the observation of Section 4 that the economists' approach to
risk aversion would lead to target stabilisation by flexible instru-
ment settings, whereas the statistical approach opts for relatively
sticky instruments.
d _ nT
HKB - I (lIA,)({3; 1( 11 ) '
,
which is the harmonic mean of the optimal generalised ridge
factors d, in (26);
(ii) the Lawless-Wang estimator,
d _ nT
LW - I ({3; 18,J ,
,
in which the denominator of the harmonic mean is replaced by
1:, A, (1 Id,*); and
(iii) the James-Stein rule, where
b+ = max {O, 1 - a*(1 - R2)1 R2}
for the optimal a* = (nT - 2)j(mT - nT + 2) and R2 is
analogous to the well-known regression measure of the same
name, applied to (8).
It can easily be seen that AI • d LW !'( dHKB !'( AnT· d LW , where )'1
and AnT are, respectively, the smallest and the largest eigenvalues
100
Lawless-Wang (d = 1.58)
Hoerl-Kennard (d = 2.06)
123.8
109.6
'"
13.1
10.7
d = 4.00 75.0 5.5
Optimal ridge regression 66.5 4.8
James-Stein rule (<5+ = 0.40) 45.0 1.8
d = 14.00 43.7 l.l
d, = 60.d,* 43.7 1.0
Ex post optimal decisions 42.4 0.0
8. Conclusion
RA
- - + - - - - - - t CE
--.-----~~~----~----~~
20-;
10 ..;
COST TO GO
1tKl-
160 ••
120 -
60 _.
,
20-:
o+----------- -~r-------~--,-----~-r--------__j
2 3 4 5
Year
Figure 2_
largest costs of adjustment (Figure I), but they do bring the ex post
outcome closer to the certainty equivalence one in comparison to
the ridge regression results (Figure 2).
Of course, both minimum variance and mean square error
decisions rely on the estimation of moments of the stochastic distri-
bution and could be improved if more accurate estimates were
available. The ridge analogy moreover faces the fundamental dif-
ficulty of assessing the true (ex post optimal) values of the instru-
ments. For that reason, a great number of ridge factors have been
proposed in the regression literature, which try to improve the
stability of the estimators. Some of these have been applied in this
paper, and they show there is scope for improvement by biasing the
instruments towards their desired values. Given that it is possible to
construct policies with smaller mean square errors than the cer-
tainty equivalent decisions, the important question remains whether
we want decision rules which are such that the overall performance
of the economy over time is expected to be robust against stochas-
tic shocks, or whether the policies themselves should be more stable
in the sense that they do not have to be changed each time the
information set is updated.
References
Arrow, K. J.: Essays in the Theory of Risk Bearing. Amsterdam: North Holland,
1970.
Brainard. W.: "Uncertainty and the Effectiveness of Policy", American Economic
Revle~\' (Papers and Proceedings) 57, 411-425.
Brandsma, A. S. and A. J. Hughes Hallett: "Dynamic Risk Sensitive Optimiz-
ation and von Neumann-Morgenstern Decision Theory", Discussion paper
8401/G, Erasmus University, Rotterdam, 1984.
Brandsma, A. S., A. J. Hughes Hallett, and N. van der Windt: "Optimal Con-
trol of Large Nonlinear Models: An Efficient Method of Policy Search Applied
to the Dutch Economy", Journal of Policy Modeling 5 (1983),253-270.
Brandsma, A. S., A. J. Hughes Hallett, and N. van der Windt: "Optimal
Economic Policies and Uncertainty: The Case against Policy Selection by Non-
linear Programming", Computers and Operations Research 11 (1984), 179-197.
Chow, C. G.: Analysis and Control of Dynamic Economic Systems. New York:
John Wiley & Sons, 1975.
Hughes Hallett, A. J.: "On Alternative Methods of Generating Risk Sensitive
Decision Rules", Economics Letters 16 (l984a), 37-44.
104
Hughes Hallett, A. J.: "The Use of Ridge Regression Techniques for Generating
Risk Sensitive Decision Rules", Communications in Statistics 13 (1984b),
127-138.
Hughes Hallett A. J. and H. J. B. Rees: Quantitative Economic Policies and Inter-
active Planning. Cambridge: Cambridge University Press, 1983.
Johansen, L.: "Parametric Certainty Equivalent Procedures in Decision Making
under Uncertainty", Zeitschriftfor Nationalokonomie 40 (1980),257-279.
Johansen, L.: LeCtures on Macroeconomic Planning, Part 2. Amsterdam: North
Holland, 1978.
Judge, G. J., W. E. Griffiths, R. C. Hill, and T. C. Lee: The Theory and Practice
of Econometrics. New York: John Wiley & Sons, 1980.
Malinvaud, E.: Lectures on Microeconomic Theory. Amsterdam: North Holland,
1972.
Newbery, D. M. G. and J. E. Stiglitz: The Theory of Commodity Price Stabiliz-
ation. Oxford: Oxford University Press, 1981.
Theil, H.: Optimal Decision Rules for Government and Industry. Amsterdam:
North Holland, 1964.
Theobald, C. M.: "Generalizations of Mean Square Error Applied to Ridge
Regression", Journal of the Royal Statistical Society, Series B, 36 (1974),
103-106.
Vinod, H. D.: "A Survey of Ridge Regression and Related Techniques for
Improvements over Ordinary Least Squares", Review of Economics and Statis-
tics 60 (1978), 121-13\.
Waud, R. M.: "Asymmetric Policymaker Utility Functions and Optimal Policy
under Uncertainty," Econometrica 44 (1976), 53-66.
Whittle, P.: Optimization over Time, Volume \. New York: John Wiley & Sons,
1982.
105
CHAPTER 7
1. Introduction
During the last 13 years the world economy was faced with floating
exchange rates of the major currencies. After a short initial period
where enthousiasm prevailed, negative effects of floating exchange
rate:>, like short-term overshooting and long-term deviations of
rates from their equilibrium levels, became visible and, as a conse-
quence, doubts about the choice of favor of floating arose. It is,
therefore, not astonishing that a renewed interest in the positive
effects of stable exchange rates can be perceived. In theoretical
studies this recent shift was witnessed by an intensified attention to
mixtures of fixed and flexible exchange rates.!
Such combinations of flexible and fixed exchange rates can be
roughly divided into three variants. These variants are charac-
terized, successively, by a crawling peg, the use of a reference rate
and an asymmetrical peg-which may be allowed to crawl. In case
of a crawling peg, the parity-provided with a fluctuation margin-
is periodically changed a little when necessary. In the meantime the
exchange rate's value has to stay within this fluctuation margin by
means of official interventions in the foreign-exchange market. To
what extent the peg or parity will be changed, depends on either a
formula chosen for this purpose or on just a policy decision. 2,3 On
the other hand, in a system based on a reference rate monetary
authorities have no commitment to intervene in the exchange mar-
ket. By contrast, they are restricted in conducting foreign-exchange
market interventions: the only policy rule is that in the event the
exchange rate is outside the fluctuation margin around the parity,
C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
©1987 Martinus NijhofJ Publishers (Kluwer), Dordrecht
106
(1)
under the constraint
Yt = A? Yt + A: Yt-l + BtX t _ 1 + CtZ t - 1, Yo gIven. (2)
In the above formulation (cf. Pindyck (1973, pp. 27 and 92)) three
types of variables are distinguished, viz. state variables YI' control
variables Xt and exogenous variables Zt. The character is used as I
Yt = MtYt-1 + (/ - M t ) Yt (3)
Y"t = L
j= I
a~+I,jYj,t + a~+I,n+IYn+l,t+1
+ L bn+l,jxj,t_1
j=1
s
L
+ j=l cn+l,jZj,t_1 + Cn+l,s+IZs+l,l-l' (5)
(7)
Vectors with a indicate the extended part of the original vectors.
Thus, the vector Yt is the vector of endogenized target variables.
Then, the control problem becomes:
(8)
with
vt =
[;:l Vt
~ [~J Qt' -- [Q, Qt~J 0 Pt [;:J
[ A't
°"J _OF,,]
nOt [ A" nI -
I _tD? -Ft t - -D)
n 2t =
[~~J .,~[C,
t -Kt
_OJ
Further, it can be noted that when all target variables are endogen-
ized the vector Yt is equal to the null vector and the matrix Qt to the
null matrix. By making use of equations (8) and (9) we can con-
struct the augmented Lagrange function (cf. Chow (1975, p. 149»:
(10)
(11)
oL (12)
OAt
By setting equations (10), (11) and (12) equal to zero and imposing
the transversality condition AT + I = 0, we can derive the optimal
values of the control variables, which we will not pursue here.
The second-order conditions can be checked easily. The bordered
Hessian matrix HI is:
116
Sufficient conditions for a minimum of the function L(.) are that the
following principal minors of the determinant H; are negative
(Chiang, 1974). This leads to the condition:
the bounds of the fluctuation margin are set at - 2.5 and + 2.5 per
cent of the par value. This bears resemblance to the fluctuation
margin of the exchange-rate arrangements of, successively, Bretton
Woods (until March 1973), the so-called Snake arrangement
(between April 1971 and March 1979) and the European Monetary
System (after March 1979).
The first crawling-peg version which we will introduce, version
2.0, is characterized by an endogenous parity that crawls according
to the relationship:
EI* - El*--I = E I - 1 - E I - 2 (version 2.0)
where E1' and EI are the exchange rate's parity and actual (or
optimal) level, respectively, in period t. Here the exchange rate is the
price of a Dutch guilder expressed in foreign currencies. In version
2.0 a change in the parity is induced by an identical change in the
actual exchange rate in the preceding period. This seems to be the
most natural exchange-rate rule. In order to emphasize that, in
practice, a change of the parity generally will be carried out by
taking into account longer-run tendencies, in the next experiment
the parity is adjusted on the basis of exchange-rate developments in
two foregoing periods. In the decision model this is identical to two
state variables. It is obvious that the most recent exchange-rate
change must dominate the effect of the exchange rate. This is
materialized by a weight for (EI _1 - E'_2) which is about twice as
large as that for (E'_2 - E. 3):
E,* - E'*--1 = 0.7(E'_1 - E'_2) + 0.3(Et _2 - Et - 3 )
(version 2.1)
Likewise, two formulae for the volume of official foreign-
exchange market interventions as a measure of parity changes are
introduced. Again one with the influence of the last two quarters:
Et* - El*--I = 0.7(RESI _ 1 - RESI _2)
+ 0.3(RESt _2 - RESt _2) (version 3.0)
and one with interventions in only the preceding quarter as the
determining variable:
(version 3.1)
where RES, stands for the level of monetary reserves at the end of
period t. As a consequence of the different dimensions of the
119
Et-2PXCt-2/PXt-2
(version 4.0)
where P XCt and P X t represent the price level of competitive exports
and the Dutch export-price level, respectively. Both price levels are
denominated in Dutch guilders.
For all the exchange-rate systems introduced in the preceding
part of this section, optimal policies have been determined. Table 1
presents for each of these policies the total costs-or loss-and,
moreover, the contributions of the time paths of the individual
objective variables to the total costs.
In the total costs reported, the de facto costs of the exchange rate
have been left out of consideration. This contributes to the achieve-
ment of a fair-minded appraisal of the systems, since it is not
relevant what exactly the time path of the exchange rate is, even if
a system has a fluctuation margin. The only purpose of the
exchange rate's weight in the loss function is to ensure that the
optimal time path of the rate stays within the fluctuation margin
around the parity. As long as this goal is pursued successfully, a real
loss in terms of economic costs is completely absent. Moreover, the
attainment of this goal can easily be judged by investigating
whether the optimal exchange rate lies outside this margin. When
this was the case, the exchange rate's weight was increased to such
a level that the optimal rate no longer crosses the margin. The costs
in economic terms of this higher weight will appear in the form of
a poorer realization of the desired time paths of the other means
and ends because of the greater emphasis on a small difference
between the optimal and the desired time paths of the exchange
rate.
In order to increase the comparability of the optimal policies,
they are subjected to some conditions pertaining to the outcome in
120
the last quarter of the period. First, it has been required that the
level of monetary reserves at the end of the period be roughly equal
to that at the beginning. Thus, it is made impossible for an
exchange-rate system to reduce total loss or costs by a net consump-
tion of monetary reserves, and in so doing burden the policy that
will be conducted in the next period. Despite this condition, the
stock of the country's international liquidity appears to grow under
all exchange-rate systems because a specified increase of the, offic-
ially held, longer-term foreign financial assets has been allowed. A
second requirement is that substantial divergences in the use of
sources for government expenditures over the entire period must be
prevented by assigning proper weights in the loss function to the
three instruments of financing government expenditures.
In Table 1 the disappointing results of the versions 3.0 and 3.1
are striking. The total costs of these versions, which employ official
interventions as the measure for par-value changes, are much
higher compared to the other exchange-rate arrangements pre-
sented in Table 1. Especially, bad records are observable with
respect to the level of real GNP and the rate of inflation. According
to Figure 1, the latter disadvantage may mainly be attributed
to the high degree of imported inflation due to the structural
depreciation of the guilder. In this respect the official-interventions
version distinguishes itself from the other versions in the diagram.
The effect suggested, is supported by a comparison of the costs
of inflation in Table 1 with the tendencies in the par values in
Figure 1. It emerges that the costs concerned rise as the trend in the
par value declines.
The remaining crawling-peg versions 2.0, 2.1 and 4.0 all lead to
lower total costs than the exchange-rate system that actually func-
tioned in the time period under examination, viz. version l.0. The
crawling-peg versions in question are based on an exchange-rate
and a purchasing-power formula. Compared to version 1.0, they
appear to ensure a high and stable growth of GNP - especially the
purchasing-power version, low inflation, and a stable terms of trade
for the export sector of the economy. A negative aspect of these
versions is the divergence between the realized and desired balances
of current international payments. In all three crawling-peg ver-
sions 2.0, 2.1 and 4.0 the guilder shows in Figure 1 a higher
appreciation when compared with version 1.0. As the relatively
strong appreciations are accompanied by relatively low total costs,
Table I. Varying parity rules and optimal-policy costs.
Exchange-rate Total costs* Contributions of the target variables
version
Real GNP Inflation Current Change terms of trade
account
Level Change Exports Imports
\.0 7549 5090 52 353 47 1 48 164
2.0 7507 5116 42 314 51 24t 144t
2.1 7370 1 5092 42 276 t 53 32 147
3.0 8234 5194 47 434 64 35 174
3.1 8246 5208 47 441 63 37 176
4.0 7403 5067 t 37t 278 52 27 194
* The total costs include both costs of the target variables and the costs of the instrumental variables.
1 Best record in the column.
tv
122
parity
150
115
exchange- rate version
liO
______ 1.0
2.0
2 1
3 0
/
135
_____ 4.0
/
130
/
125 /,"'/' /
120
/////
115
/
110
105
11 16 21 28
time period
Notes
1. See, for instance, Helpman and Razin (1979) and (1980), Lapan and Enders
(1980), Helpman (1981), Frenkel and Aizenman (1982).
2. The authors who proposed a crawling-peg exchange-rate system were in favor
of the second variant. See Williamson (198Ia), Table 1.1., p. 6.
3. As already mentioned by Swoboda (1983), from a policy point of view,
a crawling-peg exchange-rate system is familiar to a system of managed
floating.
4. Ethier and Bloomfield (1975) and (1978) deal with this system.
5. This integration plays a central role in the so-called OPTIeA proposal. See
Basevi and De Grauwe (1978), p. 145.
6. In discussions on the application of optimal control techniques, for example,
this objection has frequently been raised. See, amongst others, Prescott (1977).
7. This drawback is counteracted to a certain extent in the present study by the
fact that the model's estimation period partly concerns years characterized by
an adjustable peg, while in the rest of the period examined the guilder did not
float against all other currencies. In addition, currencies are also susceptible to
speculative attacks once parity changes become an accepted way of relieving
exchange-market pressure built up under fixed exchange rates. Both the final
years of Bretton Woods and the so-called Snake arrangement show evidence
in defense of this argument.
8. For example, in the studies mentioned in note I.
9. The complete model, including its stability characteristics, has been presented
in Jager (1981, pp. 140-156 and 180-193) and Jager (1982, pp. 233~242).
10. For details concerning the desired time paths used for the economic objectives
and instruments, see Jager (1981, 156-169 and pp. 197~206) and Jager (1982,
pp. 243~245).
125
11. The rate of return on monetary reserves and the oppertunity costs of holding
reserves are determined by the application of portfolio analysis to short-term
and long-term investments in reserve assets, respectively. Apart from interest
rates, the gains from exchange-rate of price changes have also been included
in the rates of return.
12. The following weights have been obtained. Growth of GNP: 19; stability of
GNP growth: 15; balance on current account: 7; the two terms of trade: 5. The
weight of the exchange rate was allowed to vary in order to realize an exchange
rate within the fluctuation margin. See Section 5 for further details.
13. The algorithm developed by Stoppler and Stein contains, in addition, some
minor errors due to a wrong sign in their equation (5), p. 224.
References
Basevi, G. and P. de Grauwe (1978): Vicious and virtuous circles and the OPTICA
proposal: A two-country analysis, in: M. Fratianni and T. Peeters (eds.), One
Money for Europe. Macmillan, London and Basingstoke, 144-157.
Chiang, A.-C. (1974): Fundamental Methods of Mathematical Economics.
McGraw-Hill, New York.
Chow, G. C. (1975): Analysis and Control of Dynamic Economic Systems. John
Wiley, New York.
Ethier, W. and A. I. Bloomfield (1975): Managing the Managed Float, Essays in
International Finance, 112, Princeton (N.J.).
Ethier, W. and A. I. Bloomfield (1978): The reference rate proposal and the recent
experience, Banca Nazionale del Lavoro Quarterly Review, 126, 211-232.
Frenkel, J. A. and J. Aizenman (1982): Aspects of the optimal management of
exchange rates, Journal of International Economics, 13, 231-256.
Friedlaender, A. F. (1973): Macro policy goals in the postwar period: A study in
revealed preference, Quarterly Journal of Economics, 87, 25-43.
Group of Thirty (1982): The Problem of Exchange Rates: a Policy Statement.
Group of Thirty, New York.
Hadley, G. (1974): Linear Algebra. Addison-Wesley, Massachussets.
Helpman, E. (1981): An exploration in the theory of exchange rate regimes,
Journal of Political Economy, 89, 865-890.
Helpman, E. and A. Razin (1979): Towards a consistent comparison of alternative
exchange rate regimes, Canadian Journal of Economics, 12, 394-409.
Helpman, E. and A. Razin (1980): A comparison of exchange rate regimes in the
presence of imperfect capital markets. Institute for International Studies Seminar
Paper 156. Institute for International Studies, Stockholm.
Jager, H. (1981): De behoefte aan internationale monetaire reserves als uitvloeisei
van optimaie economische politiek (The need for international reserves conse-
quent upon optimal economic policy), dissertation. Groningen, the Nether-
lands.
Jager, H. (1982): Optimal exchange-rate policy in an open economy, The Econom-
ist, 130, 228-263.
126
CHAPTER 8
C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic AnalysIS
© 1987 Martinus NiJhoff Publishers (Kluwer), Dordrecht
128
[ 11.~
I1.Kt
]
=
[0.168
° -2.297
0.951
0] [I1.Pt+l]
° I1.Kt+1 +
[-14.529]
-0.711 I1.IPt.
I1.IPt ° ° ° I1.IPt+I 1
This representation is of Chow type [2] in which the instrument
variables are incorporated in the endogenous vector in order to
simplify the objective function.
The exogenous non-controllable variables IPt and At may be
forecast in a suitable manner and equal their real values, and can
therefore be neglected in the further discussion for convenience.
The second planning sequence of periods 11 to 20 is started with
a deterministic feedback rule solving the optimization problem
backwards in time as in the first ten periods. Starting with values of
the tenth period we desire a 2.5% and a 0.5% decrease for ~ and K t
whereas IPt is held constant. For correcting the nominal trajectory
131
: 179~
1 : 241
1069~
1014j
959
"-
"-
904 ~
84 9~
794
feedforward-fltted values
629
, , , ~,--~-"--~--,,-----,,--~--,,--~-,,
o 2 4 6 8 I0 I2 I4 I6 18 20
1150
1102
1054
1006
958
910
862
814
~-~~i~
/~ / / / - - - ___ ~ F
I
/
,I
r--"T".-.----,------~,--~ , .-."---r----,'-
6 8 10 I4 16 18
::::1
1069
10 14
!
959 1
904
849
794
--,-~
I I I I I I I I
0 2 4 6 8 10 12 14 16 18
I 120
1075
1030
/
985 /
/
940 /
/
895 /
850 /
/
805 /
760
715
670
o 2 6 8 10 12 14 16 18 20
26
24
22
20
18
16
14
12
10
8
References
CHAPTER 9
Aida Montesano
University of Milan, Italy
1. Introduction
C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NlJhoff Publishers (Kluwer), Dordrecht
136
U{a} = ito
tl
u(t; a)dt. (1)
U{a} = Ito
tI
u(a(t), x(t), t)dt
with {3 > 0 and F being any increasing function, while it does not
happen in the case of non-linear transformations of u(t).
The relation between the utility functional (1) and the preference
ordering requires a deeper analysis in case the choice is made in
conditions of uncertainty. In general terms the intertemJ'oral choice
in conditions of uncertainty concerns actions which can be indi-
cated by means of distributions of probability on functions of time
which represent their consequences. In other words, by indicating
the set of consequences with C c En, actions are distributions of
probability on C, i.e., functionals a: C ~ I where I is the unitary
real interval, with
a{c} ~ 0 and I a{c} = 1.
CEe
We will consider first the case when only the functions bare
random, second the case when only the terminal time t, is random,
and third the general case when both terminal time t1 and functions
b are random.
When only the functions b are random we have the objective
functional
L
hE B
a{b; td I
tI
to
u(t; t" b)dt
(3)
where a(/l; b) is the probability that the terminal time is II. In this
scheme, set A of feasible actions, which is composed of the feasible
distributions of probability on C = TI X B, associates a distri-
bution of probability on interval TI with some points of B. Let us
now indicate the distribution of probability on TI by means of a
function of density of probability J(/I; b): i.e., the probability that
the time period finishes in the interval between II and tl + dt l is
given by J(tl) d/ l . Consequently
If function u(t; II' b) does not depend on II, i.e., it is of the kind
u(t; b), then, by introducing function
get; b) = r~ J(/I; b) dt l ,
we obtain
i.e.
Unm{a} = fCJ)
10
(L a{b} foc J(tl: b)u(t; t l, b)dt l ) dt,
bEB 1
(5)
which is again a utility functional of type (1). Ifit is u(t; b), then we
have
Unm{a} = f~
10
L
bE B
a{b}g(t: b)u(t; b)dt,
5. Conclusions
Notes
1. See, for instance, the volumes edited by Shell (1967) and by Cass and Shell
(1976). Analyses of problems of stochastic control have also been used
frequently in recent times: an introduction to them is given by Tintner and
Sengupta (1972).
2. See Richter (1971), p. 43.
3. An introduction to the mathematical theory of dynamic optimization with
reference to economic problems is given by Intriligator (1971), pp. 291--448.
4. See Montesano (1982).
References
Casso David and Karl Shell (eds), The Hamiltonian Approach to Dynamic Econ-
omics, New York: Academic Press, 1976.
Intriligator, Michael D., Mathematical Optimization and Economic Theory,
Englewood Cliffs, N.J.: Prentice-Hall, 1971.
Montesano, Aldo, "The Ordinal Utility under Uncertainty", Rivista Internazionale
di Scienze Economiche e Commerciali, 1982, 29, 442--446.
Pontryagin, L. S., V. G. Boltyanskii, R. V. Gamkrelidze and E. F. Mishchenko,
The Mathematical Theory o/Optimal Processes, New York: Interscience Pub!.,
Wiley, 1962.
Ramsey, Frank P., "A Mathematical Theory of Saving", Economic Journal, 1928,
38, 543-559.
Richter, Marcel K., "Rational Choice", in John S. Chipman, Leonid Hurwicz,
Marcel K. Richter and Hugo F. Sonnenschein (eds), Preferences, Utility, and
Demand, New York: Harcourt Brace Iovanovich, 1971, 29-58.
Shell, Karl (ed.), Essays on the Theory of Optimal Economic Growth, Cambridge:
MIT Press, 1967.
Tintner, Gerhard and Jati K. Sengupta, Stochastic Economics (Stochastic Processes,
Control, and Programming), New York: Academic Press, 1972.
143
CHAPTER 10
1. Introduction
C. Carraro and D. Sartore (eds.) Developments of Control Theory for EconomIc Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
144
2.1
The matrix Q is given by the inverse of the Hessian of the
log-likelihood function. The analytical expression of the i, jth
block of the Hessian is given in Amemiya (1977, eq. 3.5)
2.2
The matrix Q is given by the inverse of the generalized least
squares type matrix introduced in Amemiya (1977, p. 963) and
experimented with in Dagenais (1978). Such a matrix is obtained as
follows. We first introduce the T x m matrix F, whose t, ith
element is};(Yt, XI> a) = U u (the matrix of residuals) and the matrix
G" whose t th row is g;t (in practice, for models linear in the
coefficients, the matrix with the values of the explanatory variables
appearing in the ith equation). We define, now,
G, = G, - T- 1 F L (oga/ou;)' (5)
and build the block diagonal matrix G, whose m diagonal blocks are
C,. The generalized least squares type matrix used in the gradient
procedure is the inverse of the matrix
(6)
147
(for linear models, G has the form of the matrix used in Aitken-
Zellner estimation, containing the values of the explanatory vari-
ables, but with the historical values of the endogenous variables
replaced by the computed values).
2.3
The matrix Q is given by the inverse of the outer product matrix
proposed in Berndt et al. (1974) whose i, j th block is
3. Experimental Comparison
behavior did not change with the different choice of the predeter-
mined variables in the sample period (exogenous variables have
been either kept fixed in all experiments, or randomly generated
with given means and covariance matrix, and lagged endogenous
variables have also been kept fixed in all experiments, or randomly
generated using dynamic stochastic simulation), and with the dif-
ferent choice of the "true" parameters of the model, on which
Monte Carlo generations are based.
The simple computation of the number of iterations required to
get convergence with the three matrices is not particularly illumi-
nating (some more details can be found in Calzolari and Panattoni,
1983). The only sure indications which were obtained are the fol-
lowing.
(1) The use of the Hessian never requires very long tails for the
convergence, while the other two matrices (the outer product
matrix, in particular) often do.
(2) The Hessian, apart from the computational burden, rises more
often than the other problems of false convergence to saddle
points when it is used for the estimation of rather complex
models (about one out of five cases with the Klein-Goldberger
model with less than 50 observations).
Much more interesting considerations are obtained if we have a
better insight in the convergence process. For each Monte Carlo
replication, we first compute the maximum with a very high pre-
cision, then we measure the fraction of the distance between the
starting point and the maximum covered at each iteration, with the
three methods. The distance is measured both on the values of the
log-likelihood and as length of the difference between the cur-
rent and the final coefficient vectors. As before, in some cases the
two measures give different results, but the overall behavior is
practically the same. In Figure 1 results related to the distances
measured on the values of the log-likelihood function are displayed
on a log-scale. Ifwe call D(k) the distance which, after k iterations,
still remains to get to the maximum, the value which is calculated
IS
8 • 8
• •
• •
6
6 •
• • •
•• 4
• ••
4 j •
••
'I ~ *
I
:I< :I< * *
d(k) , * *
*
k= 2 4 8 10 k= 5 10 15 20
Log-linear Klein-I Klein-Goldberger
model model
8
•
B •
•
•
••
• • 6 •
6
• •
• .* ••
•• 4
•
4
•
• • •
2,.. • *
• :I<
• • :I< 2
• :I<
•• *:1< * *
:1<:1<:1<
•
••
• *:1<*:1<:1<*
**
d(k)··* *,* d(kl~_.~:1<_**~*__~__~__________
k= 5 10 15 20 5 10 15 20 25
•••
.. . Hessian
Generalized least squares type matrix
Outer product matrix
* **
Figure 1. Average rate of convergence of the three gradient algorithms.
observed for the models in Figure 1, where the length of the sample
periods are those of the historical data originally proposed for the
models themselves (only for the Klein-Goldberger model the
sample had to be enlarged of a few observations). The gradient
algorithm, which makes use of the generalized least squares type
matrix is considerably faster in the first iterations and, on average,
it allows to cover a good deal of the distance from a "good" starting
point up to the maximum (more than 99.9% for these experiments
based on rather short samples) in a smaller number of iterations
that the same algorithm which makes use of the other two matrices.
The dominance of the Hessian matrix becomes effective only in a
very tight neighborhood of the optimum, where it allows a con-
siderable reduction of the number of iterations.
8
• •
• •
6
• •
• "
• " "
• "
" "
• •
•.
4 •
••
"
"
"."
••
•
>\'
2
• •• >\' *
." • • •
>\'
•• >\' *
"
•"
.>\,. •
>\' *
>\' >\'
>\'
*
>\'
>\'
d( kJ
5 10 15 20 25
• • • • •• "Hessian
Figure 2. Average rate of convergence of the three gradient algorithms and of the
mixed gradient algorithm on the Klein-Goldberger model.
152
References
CHAPTER II
Berc Rustem
Imperial College of Science and Technology, London, UK
1. Introduction
C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NijhofJ Publishers (Kluwer), Dordrecht
158
using each individual model, to the other rival models and choosing
the strategy that causes least damage in the real world happens to
be reflected by one other than that on which the optimal policy is
based. In this section we consider the alternative when the optimal
policy is based on a combination of rival models. The way in which
this combination is defined is discussed in later sections. It is shown
that the algorithm below is general enough to apply for all such a
priori combinations of models discussed below.
Throughout this section and Sections 3 and 5 only two rival
models are assumed to exist. This is to ensure simplicity. However,
the arguments below can be trivially generalised to more than two
models. Thus, let
(2.1)
denote the first model with Y I as the vector of endogenous variables
of Fl. FI is a vector valued, possibly nonlinear, function and V is
the vector of policy instruments. Similarly, the second model is
written as
(2.2)
In general, V may be regarded as an augmented vector of the policy
instruments of FI and F2 with V ~ [vi, vrf where VI and V 2 are
respectively the policy instruments of F I and F 2. The case in which
FI is also a function ofY 2 and F2 is also a function ofY I is discussed
in Sections 6 and 7. The vector of policy instruments is assumed to
be common to all models, although the functional structure of F I
or F 2 may exclude some of these instruments. The exogenous values
that are not subject to optimization are assumed to have been
substituted into both models. Finally, the apparent static formulation
in (2.1) and (2.2) also includes dynamic models as discussed in
Rustem (1981), Rustem and Zarrop (1979, 1981)*.
Consider the optimization problem with an objective function
l(Y I , Y2' V) and the constraints (2.1) and (2.2). It is shown below
that such an objective function is sufficient to characterize all
possible a priori model combinations considered in this paper.
*The variables Yl> Y2, U, Fl and F2 WOUld, in this case, be stacked vectors of
dynamic variables. As a particular structure has not been assumed for F 1 , F 2, this
representation also covers models in which a causal dynamic structure does not
necessarily hold (e.g., rational expectations type models).
160
Thus,
min {l(Y" Y2 , U) IF, (Y" U) = 0, F 2 (Y2 , U) = O}, (2.3)
and if the objective function is a quadratic, separable in Y" Y2
and U,
l(Y, , Y2 , U) = ![Y, - YfYQYl[Y' - Yf]
+ ![Y2 - Y~Y QY2[Y 2 - Y~]
+ ![U - UdYQu[U - U d ] (2.4)
then the Gauss-Newton type optimization algorithms discussed in
Rustem and Zarrop (1979, 1981) and Rustem (1981) may easily be
modified to solve (2.3). The superscript d denotes desired values,
Qy 1 , Qy2 , Qu are {I,.symmetric matrices with Qu > 0 and Qy ,Qy2 ~ o.
In general, U d = [Uf , U~ Y where Uf and U~ refer to the desired
T T l
where the first two terms on the right hand side may be interpreted
as weighted least squares terms.
The Gauss-Newton method for solving (2.6) involves the iteration
(2.8)
starting from a specified initial value U o. The scalar IY..k ~ 0 is
chosen to ensure that G(Uk+I) ~ G(U k ) and the direction dk is
given by
dk = -Hk-1VG(U k) (2.9)
+ Qu(U - U d ) (2.12)
= 1,2. (3.2)
An immediate choice is (I., = lex, where ex" i = 1, 2 are determined
using Granger and Newbold's (1977) criterion of minimizing the
pooled forecast error. The way in which (I., are chosen is not really
important in this section. They are, however, specified prior to any
evaluation of Yin (3.1). The choice of (I., may depend on econometric
considerations. Note that "the need to pool forecasts (i.e.,
(I., E (0, 1), i = 1, 2) is prima facie evidence of a failure (of each
model) to encompass"* its rivals (see Hendry, 1983). It is shown in
Section 7 (Corollary 7.10) that, even if a model encompasses its rival
in a narrow deterministic sense, a min-max approach to choosing
an alternative pooling operation (see (7.5), (7.9) below) is only
biased towards the model that encompasses its rival. t ex, can also be
specified to reflect the policy maker's confidence in each model. A
way that avoids the difficulty of specifying (I., in this way is depen-
dent on the iterative specification of the objective function. This is
discussed in Section 5 and involves the tailoring of the objective
function to the requirements of the policy maker. In the above
+ Qu(U k - Ud)
The main difficulty with (3.6) is that (3.5) involves 1'; - }2 cross
terms (compare with (2.4». If these cross terms can be tolerated, an
even more general method for computing compromise solutions
which also does not require the specification of IXI and 1X2 is discussed
in Section 5. A simpler alternative to (3.4) is to consider the pooling
of performance measures (or objective functions) of the economy
under models I and 2. In a policy optimization framework, such as
a pooling has a number of interpretations.
Consider the vector minimization problem
min {J(YI' Y2 , U) I FI (Y I , U) = 0, F 2 (Y2 , U) O} (3.7)
where J is the two dimensional vector given by
J(Y" Y2, U) ~ [J, (Y" U), J 2 (Y 2 , U)f. (3.8)
Clearly, even if J, and J 2 are the same functions, the use of each
model produces Y, and Y2 which are different and this leads to
different objective functions (see (3.10) below). Thus, we consider
the general case with J" J 2 different. The minimization problem
(3.7) is a generalisation of the original minimization (2.3) with a
scalar objective function. Each element of (3.8) reflects a "pure"
strategy. J, (Y" U) stands for using Y, and U as the relevant policy
optimization variables. This implies that F, (Y" U) = 0 is the correct
model that reflects the behaviour of the economic system and that
it should be used in making policy decisions. Similarly, J 2 (Y 2 , U)
stands for using Y2 and U and implies that F 2 (Y2 , U) = 0 is correct.
For computational purposes, J, is assumed to be the quadratic J in
(2.4) with QY2 == 0 and Qu = QUI and J 2 is assumed to be (2.4) with
QYI == 0 and Qu == QU2·
Since the policy instrument (control) vector U has been assumed
to be common to both objective functions and models in (3.7), the
concept of a Nash solution (see Ho, 1970) is ruled out for (3.7). By
invoking Assumption (2.5), Y, and Y2 may be eliminated from
(3.7)-(3.8) to yield th~ vector minimization problem
min {G(U)IU E E nk }, (3.9)
where G(U) is given by
o (4.4)
such that
TV I = {V21V2 = arg inf [G 2(V I , O2) I O2 E 0Ji2]}·
and TV, = ¢ if the infimum is not achieved. The leader's decision
is determined by the optimization problem
inf {GI(V I , V 2 ) IV, E 0Ji1, V 2 E 0Ji2 } (4.6)
where GI (VI' V 2 ) = + 00 if V 2 E TV, = ¢. (4.7)
DEFINITION 4.7: A pair (Ur, Vn E 0Ji x 1 0Ji2 is a Stackelberg
equilibrium pair if (Vr, Vn solves (4.6). 0
Thus, in a Stackelberg equilibrium the leader chooses V I E 0Ji1 with
cost GI and the follower chooses V 2 E 0Ji2 with cost G2 •
As in the case of Nash strategies, we consider open-loop Stackel-
berg strategies which are sequentially updated with incoming new
information. A number of particular cases of interest arise accord-
ing to the relationship between the leader(s) and the follower(s) and
these are summarised below.
aG 3(V I , V 2, V 3 ) = 0
(4.9)
aV 3
(4.10)
170
The equalities in (4.10) are the necessary conditions for the optimality
of the problem
. lfG (V], V
mIll I 2,
aG 2 = 0, aV
V 3 ) I aV 2
aG3 = 0 }
3
(4.11)
and ),1' ),2 are the Lagrange multipliers associated with (4.8) and
(4.9) respectively.
~
av
[G'(V I,
V
2,
V)
3
),T aG
+,
3 (V I ,
av
V 2 , V 3 )] 0
) 3
for i = 1, 2 and j = i or 3.
= 0; i = 2, 3
a {
av, G I
~
+ )'='2 T
),)
a [
av) G 2 + ),1
T aG3 ]}
aV 3 0; i = 1, 2, 3.
conditions. Furthermore, the last case shows that the deeper the
hierarchy of the leaders gets, the higher is the required order of
derivatives of the followers' objective functions.
The fact that (XI' (X2 are no longer restricted by (XI + (X2 = 1, is not
a problem since once the values for (XI and (X2 are known G(U)'(XI
and (X2 may be redefined as
G'(U) = (5.6)
The scalars 11\ and 112 in (5.3) are selected such that the solution
of (2.3) is acceptable to the policy maker. The acceptability of the
optimal solution is now based on the broader assumption that both
models may be correct in varying degrees. Hence, the optimal
solution may capture the combined effects of both models. Thus, by
adjusting 11\ and 112 an acceptable solution to (2.3) may be sought.
An alternative approach which does not involve 11\, 112 is to seek
an acceptable solution to (2.3) by determining a joint objective
function like (2.4) for both models. This approach determines the
weighting matrices QYI' QY2' Qu in (2.4). The corresponding matrices
in (5.3) are 11\ QYI' 11 2 QY2' (11\ + 112 )Qu respectively. The scalars 11\,
112 in (5.3) can be considered to have been absorbed in the matrices
in (2.4). The approach, discussed below, determines the weighting
matrices in (2.4) that yield an acceptable solution of (2.3) on both
models and is an extension of the method discussed in Rustem and
Velupillai (1984).
STEP 0: Given Yf, Y~, U d , start with a basic set of initial weights for
Qv. I , Qv"
.
Qu' In the absence of any information we can assume
~
QYI I 0 0
- - 1 !
QC~ 0 0 (5.7)
I I
QV2
- - i i
0 I 0 Qu
r Y~'l
Yc2
U C
• (5.8)
[ ~:l
UC
+ b
(5.9)
It should be noted that (5.10) does not preserve the block diagonal
structure of (5.7) and this leads to a slightly more complicated
formulation of the Guass-Newton algorithm in Section 2. Thus, the
matrices obtained using (5.10) include general ~ - 1'; cross terms.
This yields a policy optimization formulation which is more general
than (3.5) and at no particularly greater computational complexity.
It can be shown that the above method ensures the acceptability of
the solution of (2.3). It is not possible to measure iX" iX2 in this
approach. The objective function is tailored to bring out the com-
bined effects of both models in an acceptable optimal solution.
It can be shown that the above method ensures the acceptability
of the solution (2.3). The following proposition indicates a desirable
characteristic of the method towards this end. Let, for simplicity,
the two models be linear and be given by
where
furthermore,
i = 1, ... , L, ± I}.
,~l
IX, = (7.3)
a* ~ 0, (1 - a*) ~ 0 (7.7c)
a* A] 0 (7.7d)
(1 - a*)22 0 (7.7e)
A], A2 ~ 0 (7.7f)
Note that, although J 1 and J 2 are the same functions, each model
implies a different value for the corresponding reduced cost function
°
Gp i = 1, 2. Assume that a* = then (7. 7a) represents the necessary
condition for a minimum of G2 (U) at U*. Thus, we have
min {G2 (U)} = G2 (U*) > G1(U*) ~ min {G 1(U)}
U U
Finally, the above discussion shows that the min max strategy is
not a simple pooling process, despite earlier suggestions in the
paper. If 0:* = 1 (or 0), model 1 (model 2) is taken to be the basis
of policy optimization. If 0:* E (0, l), then the optimal policy, U*,
is selected such that the policy maker's cost (or objective) function
is the same whether model 1 or model 2 turns out to represent the
economy.
In Becker et al. (1986) an approximate method for solving (7.3)
is discussed. This consists of evaluating (3.13) for various values of
0: (i.e., 0: = 0, 0.25, 0.5, 0.75, 1.0). As 0: changes, so do the values
of G1 (U) and G2 (U) at the corresponding optimal solution of(3.13).
Using curve fitting to determine the behaviour of G1 and G2 as a
function of 0:, at the solutions of (3.13), the rx corresponding to
G1 = G2 is obtained by invoking Proposition 7.6. The two models
involved being the National Institute of Economic and Social
Research and the H.M. Treasury models of the U.K. economy, the
value of 0: solving the min-max problem (7.3) was found to be 0.6
(where 0: = 1 implies total belief in the Treasury Model). In the
context of the above Corollary, the Treasury Model, which is
somewhat more complex, is built along views similar to those
reflected by the NIESR model. Thus, even if the former model can
be considered to encompass the latter in the sense of (7.8) (i.e., R1
denotes the Treasury Model), the optimal solution of 0: indicates
that the solution of (7.3), as predicted by the Corollary, is in the
range (0, 1) and not exactly at 0: = 1.
I = 1, ... , L
can be obtained within the game. The parameter vector (J" i = 1,
... , L can be determined to make the game strategy optimal.
In the single agent case; the determination of such optimal
185
9. Concluding Remarks
Acknowledgement
References
Athans, M et al. (1976): "Sequential Open-Loop Control of a Nonlinear Macro-
economic Model", in M. D. Intriligator (ed.), Frontiers of Quantitative Econ-
omics, North-Holland, Amsterdam.
Becker, R. G., Dwolatzky, B., Karakitsos, E., and Rustem, B. (1986): "The
Simultaneous use of Rival Models in Policy Optimization", The Economic
Journal, 96, 425-448.
Chow, G. C. (1979): "Effective Use of Economic Models in Macroeconomic
Policy Formulation", in S. Holly, B. Rustem, and M. Zarrop (eds.), Optimal
Control for Econometric Models, Macmillan, London.
186
Cruz, Jr., J. b. (1975): "A Survey of Nash and Stackelberg Strategies in Dynamic
Games", Annals of Social and Economic Measurement, 4(2, pp. 339-344.
Davidson, R. and MacKinnon, G. (1981): "Several Tests for Model Specification
in the Presence of Alternative Hypotheses", Econometrica 49, 781-793.
Granger, C. W. J. and Newbold, P. (1977): Forecasting Economic Time Series,
Academic Press, New York.
Hendry, D. (1983): "Econometric Evaluation of Linear Macro-Econometric
Models", Nuffield College, Oxford.
Hoel, P. G. (1947): "On the Choice of Forecasting Formulas", JASA, 605--611.
Ho, Y. C. (1970): "Differential Games, Dynamic Optimization and Generalized
Control Theory", JOTA 6 179-209.
Karakitsos, E. and Rustem, B. (1985): Optimal Fixed Rules and Simple Feedback
Laws in the Design of Economic Policy. Automatica, 21, 169-180.
Karakitsos, E. and Rustem, B. (1984): "Optimally Derived Fixed Rules and
Indicators", Journal of Economic Dynamics and Control, 8, 33-64.
Kushner, H. and Clark, D. S. (1978): Stochastic Approximation Methodsfor Con-
strained and Unconstrained Systems, Springer-Verlag, New York.
Luenberger, D. (1969): Optimization by Vector Space Methods, J. Wiley, New
York.
Miller, M. (1984): "Interdependence and Policy Coordination", paper presented
at the Thirteenth Money Study Group Conference, Brasenose College, Oxford.
Mizon, G. E. (1984): "The Encompassing Approach in Econometrics", University
of Southampton, paper presented at CEPR Workshop on the Evaluation of
Econometric Models.
Pau, L. F. (1975: "A Differential Game Among Sectors in a Macroeconomy",
Automatica 11,473-485.
Rustem, B. (1981): Projection Methods in Constrained Optimization and Appli-
cations to Optimal Policy Decisions, Springer-Verlag, Berlin.
Rustem, B. and Zarrop, M. (1981): "A Newton-type Algorithm for a Class of
N-Player Dynamic Games Using Nonlinear Econometric Models", in Janssen,
Pau, Straszak (eds.), Dynamic Modelling and Control of National Economies,
Pergamon Press, Oxford.
Rustem, B. and Zarrop, M. B. (1979): "A Newton-Type Method for the Control
of Nonlinear Econometric Models", JEDC 1, 283-300.
Rustem, B. and Zarrop, M. B., (1981): "A Class of Quasi-Newton Algorithms for
the Control of Nonlinear Econometric Models", Large Scale Systems 2,
105-111.
Rustem, B. and Velupillai, K. (1984): "On the Formalization of Political Prefer-
ences: A Contribution to the Frischian Scheme", European University Institute,
Discussion Paper 69, Florence.
Rustem, B. and Velupillai (1979): "On the Definition and Detection of Structural
Change", in M. Kohlmann and W. Vogel (eds.), Stochastic Control and Theory
and Stochastic Differential Systems, Springer-Verlag, Berlin.
Westcott, J. H., Karakitsos, E., Rustem, B., and Becker, R. (1981): "Memorandum
of Evidence on Monetary Policy to the Select Committee on the Treasury and
Civil Service", House of Commons, Third Report from the Treasury and Civil
Service Committee (/980-81): Monetary Policy, HMSO.
187
CHAPTER 12
A. J. Hughes Hallett*
University of Newcastle, UK
1. Introduction
C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
188
2. Methodology
Player i would be correct to use (7) if it is true that his opponent will
choose some value x~) irrespective of the value eventually chosen
191
for x(I); i.e., if 8x{J) j8X(I) = 0 although 8X(I) j8x(}) =1= 0. 4 But x~) will
generally be unknown because player j is meanwhile attempting to
optimise x(;) conditional on X(l). In that case x(i) must be determined
jointly with x(;) by solving (7) simultaneously with the correspond-
ing expression for x(;)*, to yield the open loop Nash equilibrium
decisions. However this solution continues to impose the restric-
tions 8x(;) j8X(I) = 0 on the rule for x(l)*, and 8X(I) j8x(;) = 0 on the
rule for x(;)*. Suboptimal decisions will result if those restrictions
are not valid. In fact these restrictions must be illegitimate since it
is inconsistent to operate a rule for X(l) * which recognises
8X(I) j8x(;) =1= 0 but assumes 8x(;) j8X(I) = 0, while at the same time
using a rule for x(;)* which allows 8x{J) j8X(I) =1= 0 but assumes
8X(I) j8x{J) = O.
There are therefore good reasons to reject the naive nature of an
open loop Nash equilibrium, and instead it would be sensible to
allow each player to anticipate the opponent's reactions to any
decision he (the first player) might make and to plan counter-
reactions accordingly. This is often done by introducing a Stackel-
berg game, in which the anticipated reactions of only one of the
player's reactions are restricted in this fashion while the other
player's reactions are determined as part of the optimisation. Our
proposal is to go one step further and leave both player's reactions
unrestricted. If both players recognise both reactions to be unrestric-
ted, then neither player attempts to dominate the other and a
Stackelberg war game is avoided.
Of course none of the restrictions, in either Nash or Stackelberg
games, need be zero restrictions. Each player might take his
opponent's policy rule (as opposed to decision values) as given, in
which case a feedback Nash equilibrium will emerge; or the policy
rule of one player could be fixed in this way (while the other rule is
determined within the optimisation) to yield a feedback Stackelberg
solution. In these cases the reaction functions of one or both players
are restricted to preassigned values, and cannot vary with the level
of the anticipated policy interventions. So the next step is to relax
that invariance restriction; and to do that, both reaction functions
must be determined as part of the decision process. Indeed, unless
a player is obliged for institutional reasons (or by lack of infor-
mation) to treat his opponent's responses as given whatever he
himself may decide, it will be necessary to remove any prior restric-
tions in order to deduce rational expectations of his opponent's
decisions and thus to avoid systematic errors in his own decisions.
192
(9)
193
where G(I) = R(I,!} + R(I,J) D(J). One obvious way of evaluating (9),
when D(l) and D(2) are unknown, is to construct a fixed point
between (D(l), D(2») and (X(l), X(2») satisfying (8). Indeed an iterative
procedure is implied; inserting trial values (D~l), D~2») into (8) auto-
matically generates new values (D~~ I, D~~ I) via (9):
- D(l)
s+ I ] [x(l) s+ I c(l)]
[F(I)
~(2)
s+ I ]
(10)
[ _ I X(2) F(2) C(2)
s+l s+ I s+ I
(11 )
Moreover a pair (D~), D~») and (X(l)*, X(2)*) satisfying (8) exists
since a Nash equilibrium exists for every nonzero sum game with
convex objectives and strategy sets. 6 If a solution exists for the open
loop Nash equilibrium, then one exists for the closed loop equi-
librium since the latter derives from objective function evaluations
which are better for at least one of the players, and probably for
both, A sensible way to search for this closed loop equilibrium
position is to start from the open loop solution (i.e., (10) with
D~) = 0, i = 1 and 2) and constrain the iterations in (10) and (11)
to generate improvements in one or both objective function evalu-
ations at each step. That ensures that the search will terminate. But,
because of its nonlinearity, (11) is not guaranteed to converge. We
therefore modify (11) by replacing D~'ll with yID~'ll + (1 - y.)D~'),
where 0 ~ Y, ~ I is a scalar chosen to force
hill" (i .e. , such that w(I)(X(I) X(2») ~
xn
W(I)(X(1)
I and x~~ I "down-
x(2)) for i = 1 and
s+ I, s+ I "'" s , s
2). Thus we search among the improvements available from the
reaction matrices which are generated by resolving (8) at each step.
An exhaustive search will ultimately lead to (X(l) *, X(2) *).
Thus, this modification of (11) may be used for various different
purposes. First, it helps overcome any convergence and uniqueness
problems in (10). To establish an equilibrium position, we must
pick that solution which yields the best objective function values for
both players simultaneously from the multiple solutions which (II)
may generate. Second it avoids problems of complex valued reac-
tions in the asymmetric case. Third it recognises that it is only
sensible for a player to alter his conjectures in a way which is Pareto
improving for both players, since otherwise the opponent simply
194
will not react in the way conjectured. One cannot expect any player
to react contrary to his own interests because that would amount
to imposing interpersonal comparisons without any compensating
bargain or side payments. Thus the conjectures must be incentive
compatible with (at least) the zero conjectures solution. In contrast,
removing incentive compatibility altogether, as van der Ploeg and
de Zeeuw (1986) do, automatically generates multiple solutions -
some of which are inferior to the zero conjectures case.
G~), and d~) at each t. The necessary steps, which are rather
Ifl), A(l)
the iteration at (11) shows that this is true for all conjectures except
those at the fixed (convergence) point satisfying (8) where the
conjectures turn out to equal the optimal reactions. These inconsis-
tencies are of course just a form of the Lucas critique of policy
making. The policies of one player, based on certain conjectured
responses by others, will actually generate reactions different from
those conjectured; and that will invalidate the original policy selec-
tion since the spillovers from those new reactions change the par-
ameters controlling the responses of the first player's targets to its
own instruments. For instance, the open loop Nash solution
generates policies for player 1 assuming that R(I,I) describes the
responses to domestic policy changes, whereas in fact, by (10) and
(11), R(I,I) + R(I) D\2) will determine those responses, Player 1 is
then obliged to modify its proposed action since decisions based on
R(I,I) will be suboptimal for the changed dynamics due to the
and e(2) = Ff2) EI/(c(2)J. The leader's targets would then behave as
jil) = (R(I,I) + R(I,2) D\2»)x(l) + c(l) + R(I,2)e(2)
X(2) *, we get
On the other hand, the leader might take the follower's policy
rule as given, say X~2) = KtYt-' + k t . The follower's reaction
matrix can be obtained by stacking and substituting out the Yt-l
terms by (3):
(15)
KT''lT + kT
4. The Model
6. The Results
tv
o
Vl
206
more actively both with loans and the discount rate, and now has
a more activist monetary policy than the US.
In summary, recognising policy interactions had led to some
convergence in the national policies. The US had dropped most
of its "supply side" stance in favour of some demand creating
measures; and the EEC is partly able to overcome the inertia of its
own policy responses as indicated by more flexible interventions,
using particularly the discount rate and money supply. Finally,
excepting the discount rate, US interventions are smaller and more
consistent while European interventions run at the same level but
have become more active.
Cooperation between the US and the EEC takes the optimal
policies a step further towards convergence. Government expen-
ditures rise faster in both economies, and the US social security cuts
are halved once again. Tax cuts are still in evidence, although less
prominent in the US. The activism of monetary policy has vanished
-particularly the sharp restrictions of money supply and interest
rates which appeared in the non-cooperative solutions. In fact both
loans to the government and the discount rates now follow constant
or steadily changing paths. This suggests that it may be important
to coordinate monetary policies. However, the convergence pattern
does show that cooperation may require individual policies to be
surrendered for the sake of concerted action, as in the expenditure
variables here.
Overall the cooperative policies call for reduced intervention in
the US, and for more consistent policies in both economies, com-
pared to the noncooperative strategies. The exceptions to this
pattern are the rise in government expenditure in the first two years,
followed by a drop in 1976-1978. European countries benefit from
this initial rise in government expenditures and from the continuous
US discount rate reductions. The induced stability provides the
opportunity to follow more consistent aggregate demand policies in
the EEC. They are much like the Cournot-Nash solution for the
first two years; they are similar to the non-cooperative solution in
1976 except for a cut in the loans to the government, and only show
an increasing discount rate in the last two years of the planning
period. This break in the concerted action over discount rate
changes in the EEC and the US is necessary in order for Europe to
attract capital imports which then compensate for the balance of
trade deterioration induced by the growth in government outlays.
209
which would yield the same objective function gains for each econ-
omy. Here the gains to cooperation for the US are equivalent to an
extra 0.47% annual GNP growth for 5 years, while for the EEC it
is "worth" an extra 1.35% annual GNP growth for 5 years.
On the face of it, therefore, the incentives to cooperate are not
very large for the US; and they are quite a lot smaller than the
incentives to exploit the game fully, or at least to allow normal
competitive forces to guide policy making. This may partly explain
the reluctance of the US to engage in cooperative policy making.
Thus, on the basis of these results, we can expect the US to make
very little effort to coordinate its policies with the EEC, while the
EEC would stress the importance of such coordination between the
industrialised economies.
8. Conclusions
Notes
I. These views are noted in Oudiz and Sachs (1984), and summarise the views
expressed by Helmut Schmidt, V. Giscard d'Estaing, Martin Feldstein, and
Laurence Klein, in the Economist between February and June 1983. See also
Basevi et at (1984), Begg (1983).
2. See Van der Ploeg (1982), Hughes Hallett and Brandsma (1983), Brandsma
and Hughes Hallett (l984a, b), Hughes Hallett (1984).
3. The problem of how to keep the evaluation of the multiplier matrices consis-
tent with the selection of instrument values X(l) and X(2) is examined by
Brandsma et al. (1984). For the purposes of this paper we will treat those
matrices as known, in order to abstract from further complications of par-
ameter uncertainty.
4. This is a passive policy according to Fair (1978).
5. See Bresnahan (1981), Holt (1985) and references therein. This method has
been extended here to a multi-target, multi-period form under uncertainty.
6. See Aubin (1979) or Friedman (1977). Bresnahan (1981) shows that a "rational"
conjectural equilibrium exists in linear-quadratic problems, although, if the
player's policy responses are asymmetric, there is no proof that the corre-
sponding reaction functions are real valued. If that should happen, it seems
reasonable to stop the iteration (11) at the real valued solution nearest the
consistent conjectures equilibrium such that no player is worse off than he
would be with any other conjectures (including zero conjectures). That prob-
lem apart, the purpose of (II) is to ensure that the computed policy responses
ax(l) /ax(J) actually equal the same values conjectured for them. Uniqueness is
more difficult. The open loop Nash equilibrium is unique in the linear-
quadratic case (Aubin (1979»; and it seems that the closed loop Nash equi-
librium may be unique for the same reasons since there is only one information
set here, covering all elements in c(l) and C(2). However, the "downhill" directed
search, introduced next, will in any case locate the joint minimum. This is not
to suggest that the convergence of (II) will be unique.
7. See Hughes Hallett (1984).
8. Bellman (1961), pp. 54-57. This time consistency property is a particular case
of the time consistency of optimal muItiperiod decisions under full informa-
tion and perfect rational expectations established by Tesfatsion (1984).
9. See Neese and Pindyck (1984).
References
paper presented to the Society of Economic Dynamics & Control, Nice, June
1984.
Van der Ploeg (1982) "Government Policy, Real Wage Resistance and the Reso-
lution of Conflict". European Economic Review 19,181-212.
Van der Ploeg, F. and A. J. de Zeeuw (1986): "Noncooperative Strategies for
Dynamic Policy Games and the Problem of Time Inconsistency: a Comment"
unpublished manuscript.
Van der Windt, N. and J. C. Siebrand (1984): "An Annual Model of the US
economy-Rasmus 2a" and "An Annual Model of the EC-Rasmus 3a", Dis-
cussion papers 8408/G and 84l1/G, Institute for Economic Research, Erasmus
University Rotterdam.
215
CHAPTER 13
Carlo Carraro
University of Venice, Italy
1. Introduction
In recent years, a new mathematical problem has been proposed in
the en'6ineering literature. The main features of this problem, called
closed-loop Stackelberg (CLS) problem, can be described in the
following way.
Suppose the control problem can be described as a game between
two decision-makers and suppose that one of the two players, called
the leader, has the power to announce his strategy first and to make
his strategy conditional on the other player's strategy. Can the
leader announce a strategy such that the follower is induced to
behave as if he were acting in the leader's interest? How can this
strategy be computed? What are the properties of the solution of the
game when such a strategy is announced by the leader? What are
the properties of the leader's optimal strategy? These questions can
be formalized into a mathematical problem and many papers have
recently tried to provide a solution to this (CLS) problem. I
The features of the CLS problem are not completely new in
economics. In fact, Chow (1981, Ch. 17) provides an algorithm
which can be used to compute the steady-state CLS solution of the
game, if the follower's strategy is included into the state vector of
the dynamic equation describing the economic system. However,
several aspects of the CLS problem, not explored by Chow (1981),
were recently emphasized in the engineering literature, where other
solutions of the problem have also been proposed. In particular, the
existence, uniqueness, and time-consistency of the CLS solution
have been analysed.
The author is grateful to Gregory Chow for helpful comments on a previous
version of this paper.
C. Carrara and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
©1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht·
216
2. Static Games
variables are related to the solution of the game in its normal form
by XI = sl(0J, i = 1, 2, where 0 1 represents the information set
available to each player. Furthermore, let us define player-i's
rational reaction set RI = R;(s), i, j = 1, 2, i =#; j, as
RJs) = {S, E SI: E[U-:CS" SJ' s)] ~ E[U-:(s;, Sf'S)] (2.1)
for all SI E SI}
Therefore R,(s) defines player-i's optimal reaction to player-j's
strategy and can be determined by solving the following problem:
min E[U-:(s" Sf' s)] (2.2)
s,
(2.5.2)
where it is important to stress that S2 is included into the leader's
information set. In other words, the CLS problem is solved if the
leader can determine a strategy S~ls such that the follower is induced
to behave as if he were minimizing the leader's loss function.
219
h(e], S2, sD = {o if S2 = s~
(2.7)
sf if S2 =I- s~
where sf must be such that
arg min E[U-;(S~IS, S2, 8)] = s~ (2.8)
S2
where the matrix P penalizes any deviation of S2 from s~. In this last
case, the CLS problem is solved by choosing a matrix P such that
(2.8) holds. In Section 4 we will discuss the advantage of choosing
a nonlinear threat function with respect to the credibility of the
leader's strategy.
A comparison between the control problem (2.3) and the CLS
problem (2.5) is straightforward. If both problems are solvable,
the solution provided by the CLS strategy is preferred by the
leader since he achieves the absolute minimum of his loss func-
tion. However, the superiority of the CLS solution is based on a
larger information set. In fact, in the control problem, the leader
announces his decision first, given his knowledge of the follower's
reaction set (i.e., the follower's loss function and the initial condit-
ions). In contrast, in the CLS problem the leader again announces
his strategy first, but he also knows the follower's actual strategy.
220
Suppose the leader tries to achieve (sf, sf). The most powerful CLS
strategy he can announce is:
sf if S2 = sf
Scls _ { (2.13)
1 - sr'P if S2 =f- sf.
Then, the following proposition can be proved:
PROOF: Obvious, since sr'P defines the most punitive (and effective)
strategy.
to adopt is S2 = S2 where
Therefore, the leader can achieve the absolute minimum of his loss
function if and only if s~ = S2.
Notice that the solution (SI' S2) is nothing more than the standard
Stackelberg solution of the game when the follower becomes the
leader and the leader becomes the follower.
Proposition 4 also implies that the interior of the inducible region
determined by a set of credible threats is empty. Define indeed the
follower's maximum loss when the CLS strategy is credible as B,
where
however, that the team solution does not belong to the inducible
regIOn, l.e.,
E[JtS(s\, s~, e)] > BmP.
If there exists a strategy sf such that
E[JtS(sf, s~, e)] ~ Bmp
the leader's best CLS solution will be
*'f
Sl 1 S2 = S2I
Scls _ {
I - s'{'P if S2 =1= s~
The leader is committed to carrying out his threats so that the
follower will choose S2 = s~. However, the leader, after having
observed S2 = s~, will choose Sl = s\ instead of the announced
· s I = S *I . 4
po1ICY
This is the time-consistency problem as it is presented in the
traditional control and rational expectations literature. 5
As proved by Proposition 3, the time-consistency problem (and
the following credibility problem) affects the CLS solution of the
game only if the inducible region does not contain the team solu-
tion. In contrast, we will show that the standard optimal control
policy is often time-inconsistent, even when the CLS policy is
time-consistent.
2.3. An example
Let the game be described by the following matrix:
Player 2
(0, 5) (2, 2) (3, 4)
Player 1 (1, 0) (2, 5) (4, 0)
(5, 3) (1, 1) (5, 4)
When the leader follows the most punitive strategy, the lowest loss
the follower can secure for himself is Bmp = 4, so that the inducible
regIOn IS:
IRmp = {(2, 1) (3, 1) (1, 2) (2, 3) (3, 2) (1, 3) (3, 3)}.
In contrast, if the leader's threats are credible, the follower can
loose only jj = 1, so that IR might be IR = {(2, 1) (3, 2) (2, 3)}.
225
if U2 = 3
However, this policy is not time-consistent since when the follower
has chosen U2 = 1, the leader has an incentive to pick Ul = 1 in
order to achieve U'I = 0, the minimum loss.
This is not the case if the team solution belongs to the inducible
region. Suppose that the element (1.1) of the matrix is replaced by
(1, 5) and the element (2, 1) by (0, 0). Then the strategy (2.15)
becomes time-consistent. Therefore, when (s\, sD E IRmp, the CLS
strategy is time consistent but, being defined by the inducible region
IRmp, it is based on a set of non-credible threats. We can conclude
that two conditions must be satisfied for a closed-loop Stackelberg
strategy to be credible:
(i) the inducible region must contain the team solution;
(ii) the leader is committed to his declared strategy. This could be
the result of a binding contract, an institutional arrangement or
the minimization of a long-term loss.
Of course, this last possibility (which might be introduced
through a reputation mechanism) cannot be explored by using
static games. Therefore, the next sections will discuss the CLS
solution for repeated and dynamic games.
3. Repeated Games
Follower
sf S2
V; - ~*, 0 o ~*- is
Leader 81
Vi - ~ , Vi - V2
smp
V; - Vi ~* - Vi
I / Vi - ~' Vi - ~
(3.5)
where inequalities (3.3) and (3.4) have been used to transform the
normal form (3.2) defined by the players' losses into the normal
form (3.5) defined by the players' payoffs. The positive quantity
Vi - is > 0 has been used to normalize the payoffs of each
player.
Furthermore, let us assume:
A.2. The CLS strategy does not satisfy condition (2.17). In other
words, the strategy that the leader wants the follower to adopt does
not coincide with .52 ,
These assumptions and the normal form (3.5) imply that the
results derived in Kreps-Wilson (1982a) can be applied to deter-
mine under what conditions the CLS solution is actually a possible
solution of the game. The following inequalities are indeed assumed
to hold:
V* - V
B. 1 > 2 2 == b > 0 by assumption A.2 and the definition
VI' - V2
of inducible region;
14.
-
~
<
-
>
Vr by assumptIOn
.
A.l.
Follower
si .52
Leader 51 a, 0 0, b
s~P I -1, b - 1
(i.e., PI = <5) and assume that both players remember the moves of
the game, as the game progresses. Therefore, we are dealing with a
game with imperfect information and perfect recall (see Kreps-
Wilson, I982a).
An equilibrium concept which is analogous to Selten's perfect
equilibrium but which takes into account the uncertainty intro-
duced into the model is the sequential equilibrium described by
Kreps-Wilson (l982b).
Therefore, we want to determine the sequential equilibrium of
the game (3.5). The function Pt is defined by the following four
conditions:
(iv) PI = <5
(i) the leader can achieve the absolute minimum of his multi-stage
loss function;
(ii) the leader's CLS strategy is time-consistent.
4. Dynamic Games
dynamic games are considered. The CLS strategy may indeed imply
a punishment from time t + lon, any time the follower does not
adopt the leader's desired strategy at time t.
The leader will therefore try to achieve the absolute minimum of
his loss function by using the optimal strategy
xt if X2t - 1 = Xit_1
Scls - { (4.3)
It - •.mp
.Alt 'f *
1 X2t-1 i=- X2t - 1
In other words, the follower will verify at any time t if his loss can
be reduced by choosing X2t i=- xit. If this is the case, the leader will
use his punitive strategy from time t + Ion. Therefore, the
sequence {B';'P; t = 0, 1, ... , T - 2} defines the inducible region
for the dynamic game (see Tolwinski, 1983). It must be emphasized
that in a deterministic setting the follower's decision at the last stage
of the game cannot be influenced by any threat, so that at the last
stage of the game no policy can be induced.
A common assumption is to exclude any follower's action at the
last stage of the game (Basar-Selbuz, 1979)8 or to impose some
restrictions on the leader's loss function (see Tolwinski, 1981).
However, these assumptions affect the effectiveness and not the
credibility of the CLS strategy. Indeed, they can be used to show
that the leader's CLS strategy is effective even in the last stage of the
game, so that the leader can achieve the absolute minimum of his
loss function. However, Selten's argument can again be used to
234
show that no threat will be carried out in the last period, so that
in all the other stages of the game the follower will choose a strat-
egy which differs from the leader's desired strategy. Furthermore,
the credibility of a CLS strategy for dynamic games is related to
the type of strategy (linear, nonlinear, continuous, etc.) which is
adopted by the leader. Let us consider, for example, the solution of
the CLS problem provided by Basar-Selbuz (1979) and Tolwinski
(1981). The Basar-Selbuz CLS strategy is defined by:
(4.7)
6. Conclusion
This paper has tried to achieve several goals: first, a new interest-
ing solution of the control problem has been presented and its main
features have been discussed. This solution, called Closed-Loop
Stackelberg, is based on an optimal announcement strategy so that
a credibility problem arises. Therefore, this paper has also shown
under what conditions the optimal announcement is credible.
Static, repeated and dynamic games have been considered.
However, several extensions of the results contained in this paper
should be provided. For instance, a general CLS solution for
236
stochastic games has not been proposed (see Chang-Ho, 1981, for
a first attempt) and the new problems arising when multi-level
games are considered have not been examined (see Luh-Chang-
Ning, 1984). Furthermore, more effective CLS strategies can be
determined when two ore more followers are introduced into the
game, so that the leader can exploit their interaction in order to
achieve his team solution (see Chang-Ho, 1983). Finally, several
problems related to the information structure of the two players
have not been considered. If, for example, the follower's strategy is
not observable by the leader, who must therefore induce the fol-
lower to reveal his actual decision, then the CLS strategy becomes
more complex and a two-sided credibility problem must be solved
(see Ho-Luh-Olsder, 1982).
Notes
References
Simaan, M. and Cruz, J. 8., Jr. (1973): "Aditional Aspects of the Stackelberg
Strategy in Nonzero Sum Games", Journal o/Optimization Theory and Applica-
tion, 613-626.
Tirole, J. (1983): "Jeux dinamiques: un guide pour I'utilisateur", Revue d'Economie
Politique 4, 550-575.
Tolwinski, B. (1981): "Closed-Loop Stackelberg Solution to a Multistage Linear-
Quadratic Game", Journal of Optimization Theory and Application, 485-501.
Tolwinski,8. (1983): "A Stackelberg Solution of Dynamic Games", IEEE Trans-
actions on Automatic Control, 85-93.
239
CHAPTER 14
1. Introduction
C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NijhofJ Publishers (Kluwer), Dordrecht
240
(i.e. if the dominant firm were to abandon the optimal policy and
revert to short run profit maximizing, it would always reduce
output, and price and entry would increase).
PROOF: Assuming that the dominant firm has been moving along
the optimal trajectory, at time t the short-run payoff is
h(q) = [a en! - blx*(t) - b2 q(t) - c]q(t), (19)
where x*(t) denotes that x(t) is on the optimal path. Maximizing
this short-run payoff h(q) leads to the myopic optimal output rule
aCt):
(20)
The long run optimal level of output q*(t), must also maximize
the Hamiltonian H at every point on the optimal path (equation
(13)). This leads to
(21)
where z*(t), the shadow price of entry, is necessarily a negative
quantity. Since k and b l are positive parameters, comparison of (18)
and (19) implies that q*(t) > qO(t). QED
where the risk aversion parameter (m) and the variance of the error
term (v) are defined in (3) and (4), and by b2 = b l + mv. Since
prices are inversely related to output, the myopic price pOet)
is higher than the long run optimal price p*(t), and we have
opo(t)jom > 0 and op*(t)jom > O. QED.
After the change of variables (14), the reduced form of the system
(9}-(l3) is:
X*(t) - (kb] + n)X*(t) - kb] Q*(t) + k(a - p e- nt ) (26)
Q*(t) (2b 2 )-] (2kbf + rb])X*(t) + [b]k + r - n]Q*(t) - A(t)
(27)
where
245
4. Conclusions
Table I
Time e -nl x*( t) e- ni z*(t) e- nl q*(t) p*(t)
0.0 1.0 5.1 140.0 5.51
10.0 10.3 68.2 144.0 5.36
20.0 15.9 111.0 149.0 5.34
30.0 19.3 158.0 134.0 5.41
40.0 21.5 204.0 129.0 5.50
50.0 23.0 249.0 123.0 5.62
60.0 24.2 282.0 117.0 5.75
70.0 25.1 335.0 112.0 5.89
80.0 26.0 377.0 106.0 6.03
90.0 26.7 418.0 101.0 6.17
100.0 27.4 458.0 95.6 6.31
110.0 28.1 498.0 90.3 6.45
120.0 28.8 538.0 85.1 6.59
130.0 29.4 576.0 79.9 6.72
140.0 30.0 614.0 74.8 6.85
150.0 30.6 652.0 69.8 6.99
160.0 31.2 689.0 64.9 7.11
170.0 31.8 725.0 60.1 7.24
180.0 32.4 761.0 55.3 7.37
190.0 33.0 797.0 50.6 7.49
200.0 33.6 831.0 45.9 7.61
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.25, c = 1.0, r = 0.15,
p = 2.0, a = 10.0, n = 0.055. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.
Table 2
Time e- nl x*(t) e- nl z*(t) e- nl q*(t) p*(t)
0.0 1.00 0.5 150.0 5.48
10.0 0.601 58.0 156.0 5.29
20.0 0.361 112.0 160.0 5.17
30.0 0.217 163.0 163.0 5.11
40.0 0.130 213.0 164.0 5.06
50.0 0.782E-Dl 262.0 165.0 5.04
60.0 0.470E-Dl 318.0 166.0 5.02
70.0 0.283E-Dl 357.0 166.0 5.01
80.0 0.170E-Dl 483.0 166.0 5.01
90.0 0.102E-Dl 449.0 166.0 5.00
100.0 0.619E-D2 494.0 167.0 5.00
110.0 0.376E-D2 539.0 167.0 5.00
120.0 0.320E-D2 584.0 167.0 5.00
130.0 0.142E-D2 627.0 167.0 5.00
140.0 0.891E-D3 617.0 167.0 5.00
150.0 0.574E-D3 614.0 167.0 5.00
160.0 1.384E-D3 757.0 167.0 5.00
170.0 0.170E-D3 799.0 167.0 5.00
180.0 0.201E-D3 841.0 167.0 5.00
190.0 0.160E-D3 882.0 167.0 5.00
200.0 0.135E-D3 923.0 167.0 5.00
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.01, c = 1.0, r = 0.05,
P = 2.0, a = 10.0, n = 0.051. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.
249
Table 3
Time e- nt x*(t) e- nl z*(t) e- nl q*(t) p*(t)
0.0 1.0 1.7 149.0 5.49
10.0 8.2 86.4 150.0 5.27
20.0 10.0 107.0 146.0 5.30
30.0 12.2 155.0 140.0 5.47
40.0 12.8 203.0 135.0 5.58
50.0 13.4 251.0 129.0 5.74
60.0 13.4 298.0 122.0 5.91
70.0 14.2 345.0 110.0 6.25
80.0 14.6 439.0 104.0 6.42
90.0 15.0 439.0 104.0 6.42
100.0 15.4 485.0 98.3 6.59
110.0 15.9 532.0 92.3 6.76
120.0 16.3 578.0 86.3 6.92
130.0 16.7 624.0 80.3 7.09
140.0 17.1 670.0 74.4 7.26
150.0 17.5 716.0 68.5 7.42
160.0 17.9 761.0 62.6 7.59
170.0 18.3 806.0 56.7 7.75
180.0 18.6 852.0 50.9 7.91
190.0 19.0 807.0 45.1 8.08
200.0 19.4 941.0 39.3 8.24
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.25, c = 1.0, r = 0.10,
p = 2.0, a = 10.0, n = 0.104. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.
Table 4
Time e- nl x*(t) e- nl z*(t) e- nl q*(t) p*(t)
0.0 1.00 2.2 149.0 5.49
10.0 4.23 55.0 153.0 5.28
20.0 5.97 98.1 154.0 5.20
30.0 6.89 134.0 154.0 5.17
40.0 7.41 165.0 153.0 5.18
50.0 7.71 192.0 153.0 5.19
60.0 7.90 216.0 152.0 5.22
70.0 8.01 236.0 151.0 5.24
80.0 8.10 254.0 150.0 5.26
90.0 8.16 270.0 149.0 5.28
100.0 8.21 284.0 148.0 5.30
110.0 8.24 296.0 148.0 5.32
120.0 8.28 307.0 147.0 5.34
130.0 8.32 316.0 147.0 5.35
140.0 8.33 325.0 146.0 5.36
150.0 8.35 332.0 146.0 5.37
160.0 8.37 338.0 146.0 5.38
170.0 8.39 344.0 145.0 5.39
180.0 8.40 349.0 145.0 5.40
190.0 8.41 354.0 145.0 5.40
200.0 8.42 358.0 145.0 5.41
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.10, c = 1.0, r = 0.05,
P = 2.0, a = 10.0, n = 0.064. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.
250
References
Derivation of 15 and 16
From (14);
x (0.1)
(X + nX) e r (0.2)
From (13):
o. (0.5)
251
Using (14);
(a - 2b 2Q - b]X - c e- nt - b]kZ) e(n-r)t o (0.6)
Q = (2b z)-](a - b]X - b]kZ - ce- nt ). (0.7)
Substitution of (A7) into (A4) yields:
X = ka - (kb] + n)X - (kb] /2b z)
x (a - b]X - b]kZ - c e- nt ) - kp e- nt (0.8)
X - (kb] + n - k(bf/2b 2 ))X + e(bf/2b z)Z
+ k(a - (ab] /2b 2 ) - (p + b] c/2b 2 )) e- nt • (0.9)
Using Ao and A] as described below (16) yields equation (15).
Using (14) in (11);
(2 + nZ) e nt = b] Q e nt + (b] k + r)Z e nt (0.10)
2 = (b] k + r- n)Z + b] Q. (0.11 )
Using (0.7) in (0.11);
2 = (b] k + r - n - (b] k/2b 2 ))Z - (bf/2b z )X
(0.12)
+ (b] /2b 2 )(a - c e- nt ).
Using Ao and A2 as described below (16) yields equation (16).
A? + A(2n - r)
(0.15)
Hence equation (18). If b, = b2 , then Ao = (lj2)b, and using this
result in (18) leads to;
Hence (32).
Derivation of 26 and 27
From (13), we have 8Hj8q = 0, hence
(0.17)
(0.18)
and taking the derivative with respect to time yields
z = (b, k)-' (na ent - bl X - 2bA). (0.19)
Using (14) in (2);
(X + nX) ent = k(a ent - bl Q enl - b, X e t - p) (0.20)
(
-(kb] + n) -kb] )(X\
(2kbi + rbd 2b 2 (kb] + r - n) Q}
M __ (kb/2 + n - A k 2b/2 )
(0.33)
b/2 -(kb/2) + r - n) - A
After some manipulations;
A2 + (2n - r)A + «- bkr/2) - nr + n2) o (0.34)
which has roots:
- n + (r/2) + (1/2)(r2 + 2bkr)I/2 (0.35)
- n + (r/2) - (1/2)(r2 + 2bkr)I/2. (0.36)
Assuming a solution of the form X = C eAt, Z = KC e)·t, a non-
trivial particular solution must satisfy
(0.37)
1000
100
10
Plot 1
258
100
10
Plot 2
259
1000
100
10
Plot 3
260
1000
100
10
20 40 60 80
Plot 4
261
CHAPTER 15
Klaus Conrad
University of Mannheim. FRG
1. Introduction
C. Carrara and D. Sartore (eds.) Developments of Control Theory for EconomIc AnalysIS
© 1987 Martmus Nijhoff Publishers (Kluwer). Dordrecht·
262
goodwill, the producer will loose goodwill on the market and his
demand function will shift downwards. The objective of the paper
is to show that reputation as an endogenous signal of quality can
prevent the kind of market failure characterized by Akerlof (1970)
for the case of asymmetric information on quality.
In the next section we will begin with a monopolistic industry
where the monopolist chooses price and quality under asymmetric
information starting with a given goodwill. We will characterize the
optimal paths of price and quality towards the steady state solution
and will investigate the direction of the goodwill formation process.
In the third section we will analyse the duopolistic market by
looking for a consistent solution on that market compatible with
the actions of both firms. In the fourth section we compare prices
and quality levels in the duopolistic market under perfect and under
asymmetric information.
max
p(t),n(t)
rx e-
Jo
rt {px(p, G) - C(x(p, G), n)} dt (1)
o => - Cn + px = 0 (4)
A =>fl- rp = -xG(p - cJ
+ n~xp - p(n - ne)xG' (5)
From (4) we see that the current (non-discounted) shadow price p
of goodwill is equal to the marginal cost of quality per unit sold.
A marginal increase in quality raises production costs per unit
which is the price for an additional unit of goodwill. If the quality
264
(10)
8n*
-> 0 (11 )
8J1
8n*
8G = o. (12)
dJ11 8Cj8G
dG G~O = - 8Gj8ji > 0 (see (15)). (17)
G=O
jl=O
L-______________ ~ ____________ ~G
G*
Figure I illustrates a phase diagram for (13) and (14) in the neigh-
bourhood of the steady state. Under the assumption made there is
one equilibrium which is a saddlepoint.
If Go, the initial stock of goodwill is less than G, then the optimal
path is characterized by a monotonically increasing G and a
monotonically decreasing p. The firm accumulates goodwill over
time by giving the consumers a pleasant surprise. As the state
variable G is a monotone function of time along a path converging
·to a particular steady state, G does not oscillate and producers
either permanently accumulate or decumulate goodwill.
We finally study the monotonicity of the optimal paths of p, n
and x. We find:
jJ
op*.
-G+-p
op*. {>
0 if Go < G
(19)
oG op ~ 0 if Go > G
(+) :>
on*. on*
-G+-/1
0 if Go {< <G
(20)
oG op > 0 if Go >G
(0) (+)
Let us consider the more realistic case that at the beginning Go is
lower than the steady state goodwill G. Then G increases (G > 0)
and p decreases (/1 < 0) which implies jJ > 0, price increases, and
Ii < 0, quality decreases. Thus, the monopolist offers high quality
at low prices at the beginning to accumulate goodwill. As time goes
on he reduces quality and increases the price. The quality is still
267
worth its money but one has also to pay for the reputation. With
respect to quantity, we get:
x = xpfJ + xGG.
(-) (+)
The price increase reduces the quantity but the accumulation of
goodwill can more than offset this reduction in quantity.
max
p"n,
f: e-r,1 {PIXI(PI, G I ; P2, G2) (21)
max
P2,n2
r' e- r21 {P2 X2(P2, G2; PI' G I ) - Cix 2(·), n2)} dt (22)
-HG,
(27)
n, = n"(G,) (30)
(31 )
with p, = Cu, ix, where PJ, are the opportunity costs of goodwill in
terms of producing the quality expected to sustain goodwill. As the
steady state system (25)-(28) is recursive with respect to 11" the
steady state version of (28) has no feedback on the solution set.
From (28) we obtain
(32)
270
~----~--------~I----~--------------. P,
P,2
n,
n, n, n,
L---',:--------....J2'::-----:f----------.
where G = ne-' (n); i.e., goodwill is derived from the inverse of the
former steady state conditions n = ne(G). Necessary conditions
are:
x(p, G) + pXp - Cxxp = 0 (37)
pXc - C,x c - Cnn~ = 0 (38)
as
dG
dn nec
If r = 0 in the steady state condition (7), the perfect information
solution (p, ii) and the steady state solution (ft, n) are the same. If
we assume r > 0 and compare the perfect information solution ii
of (38) with the asymmetric information solution n of (7),
(7)
we know from the second order conditions that the left hand side
of (7) is monotone decreasing in G so that G in (38) must be higher
than {; in (7) and therefore ii > n. This line of reasoning is correct
if the left of (7) is increasing in p; if it is decreasing in p, a price
increase with a lowering in goodwill is also a strategy to converge
for r ~ 0 from the solution (n, p) of (6) and (7) to the solution
(ii, p) of (37) and (38) (see also Shapiro (1982».
For a proof we have to show that the quality Fz which solves the
steady state first-order condition of the problem
m:x f e- rl
{px(p, G) - C(xC), n)} dt
6. Conclusion
Appendix
+ J1nxGG < 0,
275
and HpG > 0 if xpG < 0 does not offset all positive terms in the
following expression for HpG:
HpG = XG + xpG(p - CJ - CxxxGxp
+ Il(n - ne)xpG - Iln~xp.
an* I
-D (-x)Hpp > 0
all
an*
O. QED
aG
Proof of (15)
ac aG - nGe ) X
( an*
p*
+ (n* - ne) (xpaaG + XG ) < 0
aG
in the neighbourhood of the steady state and due to (12).
aG an* ap*
all = all x + (n - ne)xp all > 0 (see (11)).
_ H GG _ H Gp (C nn H pG ) =
-Hpp Cnn
Proof of statement 4
The first-order conditions is equation (7), i.e., with p = p and
n = n"(G):
pXG(p, G) - C.(x(·), n) - CnC)n~ = rCn/x. (7')
With G = ne - 1(n) the left-hand side is strictly monotone decreasing
in n. Under perfect information the right-hand side of (7') is zero
(see (38». Thus a solution of (7') for n implies Ii < n.
Notes
References
Akerlof, G. (1970): "The Market for Lemons: Qualitative Uncertainty and the
Market Mechanism", Quarterly Journal of Economics, 84, 488-500.
Conrad, K. (1982): "Advertising, Quality and Informationally Consistent Prices",
Zeitschriji fur die gesamte Staatswissenschaft, 138, 680-694.
Conrad, K. (1985): "Quality, Advertising and the Formation of Goodwill under
Dynamic Conditions", in G. Feichtinger (ed.), Optimal Control Theory and
Economic Analysis 2. North-Holland, Amsterdam, 215-234.
Nerlove, M. and K. H. Arrow (1962): "Optimal Advertising Policy under Dynamic
Conditions", Economica, 124-142.
Shapiro, D. (1982): "Consumer Information, Product Quality and Sellers Repu-
tation", The Bell Journal of Economics, 13, 20-35.
Thepot, J. (1983): "Marketing and Investment Policies of Duopolists in a Growing
Industry, Journal of Economic Dynamics and Control, 5, 325-358.
PART IV
CHAPTER 16
Charles S. Tapiero
Hebrew University, Israel and Case Western Reserve University
1. Introduction
C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
280
Active Passive
Risk Managarent Risk Managarent
loss
prevention Probabilities Insurance
reinsurance
risk sharing
Information
Figure 2.
282
Dynamic system
{At} -+ {Ar<1' LI' VI' ~t}
current past periods periods periods
state states lia bili ties intervention disturbance
284
A L U ~
bonds -premium assets conversions benefit payments
cash -income dividends unforseen inflation
investments benefits premium rates transfer costs
real estate expenses reim bursemen ts
which is the integral (the sum) of the measured (or expected util-
ities) benefits obtained by the insurance firm as long as it is solvent
less its costs (or disutility), less the cost of insolvency of the firm,
discounted at an appropriate rate of discount and "," is the firm's
random time reached when the firm becomes insolvent. Basically,
this time is a function of the firm position regarding its assets and
liabilities holdings. Alternatively, a function G(A, L) could be
defined such that insolvency occurs at the first instant of time
reached when.,
, E G(A, L)
In other words, when the firm reaches certain levels in the asset-
liabilities plan it may be dissolved and costs, a function of the firm's
penalities of involvency, are incurred.
In summary, then, the insurance firm would seek to
Optimize the
Temporal Objective
Including Bankruptcy Cost
Subject to
A State-Space Representation
of the Insurance Firm
+
Solvency & Regulation
Constraint
+
Other Constraints
286
To Obtain
Management Strategies for Investment
and Liquidity Maintenance
+
Premium Rates Policies Sensitive or
Insensitive to Change and
the Firm's States
+
Reserves for
Unforeseen Contingencies
and Assets Depreciations
+
Learning Mechanism for
Adaptation to Change
3. Selected Applications
Investment 1 - _...
1---...... K Capital Assets
Desinves tmen t g The Insurance
Firm
~
Premium
Dividend P ---tL..--.,:-_ _ _ _ _r---Jt------ M Liquid Assets
D --001. Cash
t
Income from
f
Claims
Assets rK F;(dt).
Hiring Quit
H sL rate
! t
L wages
Labor/
Agents
~L + aNP - Dividends DC t)
Outlays
Claims
~(dt)
Income
PremiUlll I Assets
Cash M
Contracts ~
lnves f'.
NP K
N
t Returns rK
I Income
Assets Conversion g
K = Assets ($)
rK return on assets ($/time)
g ~ assets conversion ($/time)
M .. Cash ($)
N - Contracts (Number) f; ~ claims (random)($/time)
L - Labor (Agents)
w = wage per agent ($/agent/time)
a a Commissions rates
P - Premium per contract
I a Investments ($/time)
5 = natural quit rate (a percentage)
D a Dividends ($/time)
H = Hiring or firing (Agents/time)
In this formulation, the costs of bankruptcy at time 'l: (the first time,
no more funds are available to meet claims) are given by B, dis-
counted at the rate i. The policy set states that the investment rate
can be at most equal to a fixed proportion Ie of cash holdings, that
desinvestment can at most be equal to a fixed proportion ge of
capital assets K and that the dividends rate is constrained by
[0, Dmax]. The cost of investments and desinvestments ar deducted
directly from cash with (13 - 1)1 equalling the investment trans-
action cost. Similarly, if g are desinvested then only (1 - y)g
dollars are actually added to cash holdings with g equalling the
transaction cost. Additional administrative costs, drawn out of
cash include the agents cost wL where w is the wage rate and L is
the number of agents hours per unit time. It is possible to replace
(or complement) these costs by commission costs which equal a
proportion of premium income. In this case, if rx is the commission
rate, then instead of the premium income rate PN we will have
(1 - rx)PN. The relationship between the number of insurance
contracts and the agents work force L is assumed to be given by a
Cobb-Douglas production function, such that N(t) = Nop-a L b ,
13 > 0, b > O. For simplicity, L is assumed constant (with Hiring
Rate = Quit Rate in Figure 5).
We also let bK in the "dK" equation be capital assets depreci-
ation, the rate of return r on capital assets be constant and finally
the claim rate be defined either as Weiner Process [20], [31], or by
a Jump process ([9], [32]). If the claim rate can be represented by a
Weiner process, then this means that
aJ aJ
J(K + dK, M + dM) J(K, M) + aK dK + aM dM
+ 1/2 ~; (dM)2
Note that (dK) is of order dt and therefore all terms of order (dK)2
are of the negligible order 0(dt 2). However, (dM) is of order dt and
(dM)2 has a term of order dt since (dw)2- t he Weiner component
has an expectation of order dt. With this information in hand, it is
easily proved that an equivalent formulation of our problem can be
written (using Bellman's Principle and its functional development)
as follows:
aJ (J2Na 2 J}
x aM + 2aM2
subject to
and at bankruptcy, when the insurance firm has depleted its assets,
J(K,O) = - B.
The solution of such (partial differential equations) is very difficult.
Nevertheless, for some cases, analytical results might be found,
while in others the application of numerical optimization tech-
niques (as in [34]) can be used to obtain practical results. In our
case, due to the linearity of the policy instruments I, g and D it has
been shown in [34] that a solution can be characterized by "barrier
policies" in the (K, M) plan, as shown in Figure 5. Thus, given
Premium P and any (K, M) combination, the insurance firm can
reach an optimal action regarding its policy instruments (I, g, D) by
inspection of the Figure 5.
293
.. ~3
"-
,
~
, (v)
" . '.
" :O~
" (ii)
'"
>
>
, , .~
-
I ,
-
,\
- °2
j
, "' , ). , ,
} , ) ~ (iii) \
:----v. ----1f
1°1
4 :. . 6
I
(i) (iv) I
.. I
M* M*2 M*
3 1
Policies:
(v) Invest
(1) + (11) Dc~lnve8t
x !l!...
aM + A r" [J(K, M + z) - J(K, M)] dFtz)}
subject to
o~ I ~ ~ M, 0 ~ g ~ g, K, 0 ~ D ~ Dmax.
and at bankruptcy we have the condition J(K, 0) - B stated
earlier.
294
: if z < Q
z = {
.: if z ~ Q
then a claim's expected cost will be given by:
A foX! (z - Q) dF(z).
Claims
2
o dt)
Premiums
NPdt
-x Wealth
r discount rate
~ mean claim rate
0 2 claim rate variance
N number of insured
P premium payment
x firm cash-assets
C administration costs
h opportunity cost of money
K bankruptcy cost
dw a standard Wiener process
T
Min J ex) E{fe -rt pdt + Ke- rT }
PEP 0
subject to
dx = NP L - ~N - hx - CJdt Nodw
xo > 0
Figure 6.
there is no bankruptcy;
J(x) = Min E{P dt + (1 - r dt)J(x + dx)}
PEf!J>
while at bankruptcy; x o and J(O) K.
297
Insert into J(x + dx) the stochastic wealth equation for dx and
approximate it by a two terms Taylor series expansion, or
aJ a2 J
EJ(x + dx) '" J(x) + ax E(dx) + 1/2 ar E(dr)
Since E(dx) = [NP - IlN - hx - C] dt and var (dx) = (Nu 2 /2) dt,
insertion of these equalities into Bellman's definition of optimality,
yields:
. { dJ Nu 2 d2 J
Mm - r J(x) - [NP - IlN - hx - C] dx + 2 dx2 + P
}
= 0
where
n(x) = exp [ - Rx]
a(x, I) = PN - hN - C - Q(()
where Q(() is the premium paid by the mutual firm for a (-excess
loss reinsurance contract. This means that our mutual insurance
firm covers all claims up to e while the excess claim is covered by the
reinsurer (a government for example). As stated previously, the
function Jl(/, A) is a function denoting the number of jumps of the
process X(/) in (0, I] and Jl(dv, dz) is a measure on R+ x R defined
as follows
Jl(t1., A) = Jl(t + M, A) - Jl(/, A)
and A is a Borel subset on R. The measure Jl(t1., A) is called the jump
measure of the process {X(/), I ~ O}.
For a Poisson stationary claim rate qN and claim (severity)
distribution function F(z), an application of Bellman's Principle
yields
L(s) aK4>(s)
.
+p 4>(s)
s
1
4>(s)
as + qN { 1 - f: e- ZS dF(z) + feoo e-~ dFtz) } + r
a = PN - C _ QO«().
Evidently, at the limit x --+ 0, lims oo sL(s) = K while at the limit
-+
300
then
1
¢(z)
as + r + qN{s[I - e-(S+Jl)I;]/(11 + s)}
and so
¢(s) = [l/(qN~ + a)][r/(qN~ + a)) + sri
which yields
1 e(rlNql; + a)x
qN~ + a
a = PN - C - Q - Q(~)
301
J-l +s
cf>(s) = (as + r)(J-l + s) + qNs
which yields by analytical inversion
(s\ - J-l) e -SIX - (S2 - J-l) e -S2 X
Evidently, when (Nt, rt, pi) then et and when ~t then e may
increase or decrease, depending upon the cost Q(~) of reinsurance
to the mutual insurer.
Acknowledgements
References
CHAPTER 17
James H. Gapinski
Florida State University, USA
1. Introduction
C. Carrara and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-
306
2. A Macro Model
to scale prevail, 0 < f3* < 1, and the possibilities for factor sub-
stitution exist within bounds: 0 < 0"* < 00, where 0"*, the ex ante
elasticity of substitution, equals 1/(1 + Q*). Thus, - 1 < Q* < 00.
Moreover, 0 < y* < 00 while ~i > 0, ~L > 0, and ~i + ~L = 1.
The Yv* and L! may be termed respectively the capacity output and
the capacity employment of vintage v capital.
Once extant vintage v ordinarily encounters reduced substitution
possibilities, and those which remain are summarized by the ex post
production function
v
.1 vI -- Y e /l'V[J:'01 I-a
vI + J: L-a]-P'/a
'0 Lvt . (2)
(6)
309
(9)
Price variable qi in equation (9) is real user cost.
Optimum marginal products (7) and (8) and production function
(1) can be used to express capacity labor, investment, and capacity
output in terms of the prices wi and qi. Although awkward, the
resulting expressions can be written neatly as
L*I L*(wi, qn, (10)
I(wi, qn, (11)
2.4. Prices
As argued in conjunction with equation (4), new capital is designed
under the presumption that workers bargain the money wage to
avert any erosion of the real wage. Workers do try to preserve reals,
and therefore
Xt = Pt, (23)
Xt standing for the money wage at t. By virtue of the equality
wt = xt - Pt( wt/w t_ I ), it follows that
(24)
An equation for the real wage was already introduced. Relation
(18) ties Y, and W t together for any T w , but with Y, set by statement
(19), it yields W t • Formula (18), then, is the real wage equation.
Nominal interest rate it emerges from equilibrium in the money
market. The demand for real balances is
Mdt/pt = Va + VI 1'; + v2i t(1 - Ty) + v3Mdt _ l /pt_l, (25)
where Mdt symbolizes the nominal demand for money at t and
where the nominal after-tax interest rate enters to measure the
opportunity cost of holding money. Nominal money supply at t,
Ms t , follows the feedback rule
MS t = t/to + t/t1(Ut- 1 - 0)+ + t/t2(Pt-1 - p)+. (26)
(29)
Forces controlling labor supply keep this rate from falling below
a value ~.
Machine utilization is reckoned from the capacity outputs of the
installed vintages; namely,
Y o· = .90
1550
1500
1450
1400
1350 0=.30
Figure 1.
for virtual clay, they take the shape of decided explosions? This
propensity of smaller (J to generate greater volatility reflects the
structural relationships noted in connection with equation (18). A
slight shock to Y has greater impact on w for smaller (J, and that
greater impact reverberates throughout the entire system.
Table 2 looks more closely at the volatility issue. Corroborating
the movements displayed in Figure 1, it indicates for a (J* of 0.90
that the variance of Y rises as (J begins its fall from putty and that
it soars as (J declines past the middle grade of tin, 0.60. The varia-
bility relative for Y, simply the Y variance expressed relative to its
(J = 0.90 level, drives home the point: For (J = 0.30 the variance
Y
0°= 90
1550
0=.30
::\\
.. \
\\
1500
~\
\ ,
:1
:1
:\
'.\
\\
1450 \\
\1
~\
\\
\\ +, = 685.00
'1', = 711.84
1400
+, = 685.00
,-
:'1', = 111.84
\. / +, = 85.00
-- '1', = 711.84
+, = 85.00
1350 '1', = 111.84
Figure 2.
1jI1
0
. =.90
Figure 3.
confines itself to the case of virtual clay, Since the feedbacks remain
dormant over the years immediately following the C/>O shock, output
then exhibits the same temporal motion for any c/>] and 1/11' Trigger-
ing occurs soon after 1995, and the tempering effect begins to take
hold. It becomes especially noticeable after 2005. To avoid clutter,
Figure 2 suppresses for the midterm the curves corresponding to
single policies; if drawn, they would lie between the two illustrated.
Figure 3 provides a fuller portrayal of macro behavior under the
policies. Specifically it sketches isovars, each being a curve that
shows the various combinations of the feedback coefficients necess-
ary to achieve a given degree of countercyclical control. Control is
reckoned in terms of variability composite 2, which is labeled S for
convenience. 5 At the origin S = 1.00 independently of (T,
Apart from indicating that isovars are hardly simple displace-
ments of each other, Figure 3 confirms several properties noted
322
previously. First, <PI is stronger medicine than is 1/11: For any (J the
dosage needed to work a prescribed amount of cure is less than that
ofl/l l . With (J = 0.40, for instance, as of 0.93 requires that <PI rise
by 100 units to 185.00, but it obliges that 1/1 I rise by 400 units to
511.84. Second, 1/11 has a destabilizing effect near the middle grade
of tin. When (J = 0.50 increasing 1/1 I first evokes a weakly favorable
response, but later the exacerbating tendency of 1/11 becomes domi-
nant. Maintaining the same S in that situation requires a furthering
of action from <PI and imparts a positive slope to the corresponding
isovars.6 For (J closer to clay, however, 1/11 always has a tempering
influence, and hence it and <PI are substitutes as the four negatively
sloped contours attest. Third, smaller (J mean stronger countercycli-
cal effects. As (J declines the macro system becomes more responsive
to the feedback parameters, and therefore a particular level of
control can be attained through less aggressive policy action. The
leftward shift in isovars for falling (J illustrates the case.
Section 3 reported that purchase policy is inclined to produce a
trade-off between inflation and unemployment. To see how feed-
back policy influences this relationship, the expression Pr = (Xo +
(XI / Ur is estimated within the simulation program by ordinary least
squares using the 31 observations generated from 1985 to 2015.
Intercept (xo represents the minimum rate of inflation: the asymp-
totic floor of the Phillips curve. Ratio - (XI /(Xo sites the price stability
point: the rate of unemployment at which price remains unchanged.
Both quantities, along with (XI' are scale-adjusted in Table 4, which
lists the magnitude of feedback policy as incremental from the
stationary-state values. A 300.00 in the second column means
<PI = 385.00 for <PI policy alone, 1/11 = 411.84 for 1/11 policy alone,
and <PI = 385.00 and 1/11 = 411.84 for both policies together.
Given the rudimentary nature of the regression, Table 4 remains
silent about the econometric properties of the fits.
On balance intensifying countercyclical policy rotates the Phillips
curve counterclockwise; that is, increased policy magnitude tends to
raise the minimum inflation rate, flatten the slope, and advance the
price stability point. In accord with the variability results, the most
pronounced rotation occurs by combining a (J of 0.30 with the
policies applied jointly. Then, as the control coefficients expand
over their grid of values, (Xo rises from - 0.0196 to - 0.0079, (XI
declines from 0.000490 to 0.000255, and - (XI /(Xo increases from
0.0249 to 0.0324. By contrast, and again in accord with the earlier
Table 4. Phillips curve characteristics under different types and magnitudes of feedback policy:
(J Policy Policy type
magnitude:
<PI policy alone 1/11 policy alone <PI and 1/11 policies jointly
base plus
lOao 10001X 1 - lOa l /ao lOao 1000a l - lOal/ao 101X0 1000al - lOa,/ao
0.50 0.00 -0.169 0.762 0.450 -0.169 0.762 0.450 -0.169 0.762 0.450
0.50 100.00 -0.168 0.758 0.450 -0.170 0.764 0.450 -0.168 0.759 0.451
0.50 200.00 - 0.167 0.753 0.451 -0.170 0.765 0.450 -0.167 0.755 0.452
0.50 300.00 -0.166 0.749 0.452 -0.170 0.765 0.451 -0.166 0.751 0.453
0.50 400.00 -0.165 0.744 0.452 -0.170 0.766 0.451 -0.165 0.748 0.453
0.50 500.00 -0.163 0.740 0.453 -0.170 0.767 0.451 -0.164 0.745 0.454
0.50 600.00 -0.162 0.737 0.454 -0.170 0.767 0.452 -0.166 0.757 0.455
0.40 0.00 -0.141 0.573 0.407 -0.141 0.573 0.407 -0.141 0.573 0.407
0.40 100.00 -0.138 0.568 0.411 -0.142 0.580 0.409 -0.138 0.568 0.412
0.40 200.00 -0.135 0.557 0.413 -0.141 0.580 0.411 -0.135 0.562 0.418
0.40 300.00 -0.131 0.546 0.416 -0.140 0.579 0.412 - 0.131 0.553 0.422
0.40 400.00 -0.129 0.542 0.420 -0.141 0.585 0.415 -0.128 0.544 0.427
0.40 500.00 - 0.126 0.534 0.423 -0.140 0.585 0.417 -0.125 0.538 0.431
0.40 600.00 -0.123 0.526 0.426 -0.140 0.585 0.419 -0.123 0.533 0.435
0.30 0.00 -0.196 0.490 0.249 -0.196 0.490 0.249 -0.196 0.490 0.249
0.30 100.00 -0.189 0.493 0.261 -0.198 0.508 0.257 -0.180 0.476 0.264
0.30 200.00 -0.172 0.455 0.265 -0.189 0.497 0.262 -0.151 0.399 0.265
0.30 300.00 -0.153 0.407 0.267 -0.180 0.473 0.263 -0.119 0.314 0.264
0.30 400.00 -0.133 0.355 0.267 -0.169 0.445 0.263 -0.108 0.311 0.289
0.30 500.00 -0.113 0.303 0.268 -0.158 0.414 0.262 -0.092 0.282 0.305
0.30 600.00 -0.105 0.295 0.281 -0.146 0.381 0.261 -0.079 0.255 0.324 w
N
w
.0'* = 0.90.
324
5. Concluding Comment
Capital bears on feedback control. The more dissimilar is capital
ex post versus ex ante, then the more unruly is business cycle
activity, and the more frequently and forcefully is a feedback
triggered. Capital akin to putty-clay causes marked oscillations
and therefore provides a solid opportunity for feedback success.
Conversely, putty-putty generates mild fluctuations, for which
controls are not particularly relevant.
In the capital-theoretic context, feedbacks can improve macro
behavior. They can worsen it. Nevertheless, behavior becomes
more responsive to control as capital ex post tends toward clay, and
near clay the prospect of improving performance and of doing so
appreciably is quite strong while that of worsening it is rather weak.
In fact, near clay feedback policy, including the monetary initiative,
neither deteriorates the overall behavior of the system nor provokes
serious stagflation movements even when the magnitude of policy
increases six- or eightfold. By virtue of these findings, the familiar
propositions, advanced with little regard to the nature of capital,
that active countercyclical maneuvers can make matters worse and
that policy should be applied lockstep without feedbacks may need
to be substantially revised.
Notes
l. It should be recalled at the outset that rational expectations need not disable
feedback policy. On the contrary, feedbacks can flourish under rationality as
Phelps and Taylor (1977), Chow (1981, pp. 225-238), and Howitt (1981) show.
Sargent and Wallace (1976) describe the standard rational-expectations
position.
325
References
CHAPTER 18
1. F. Richard
e.O.R.E, Louvain, Belgium
Introduction
C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
328
P
D(Xi:) I X T , Be) = n D(xT+p I XT+p-l, BJ
p=1
(2)
LI«(}ls; XT) IX nT
t=1
D(Yt I zl' Xr-I' (}Is) (4)
the sort of setup which seems to be appropriate for the analysis and
simulation of policy measures. This approach is often contrasted
with that known as "multivariate time series modelling" whereby it
is the joint process D(x t I X;-I' .) which is modelled directly in
terms of its "reduced" or "final" forms parametrisations. In the
present context the corresponding parameters are functions of both
() Is and ()2s'
Up to the use and testing of theory based prior information, the
choices between these two modelling approaches might not prove
critical in worlds where parameters do not change. In particular, the
time series modelling approach then provides us with a parametri-
sation which is convenient for investigating the Granger causality
structure of a multivariate process as in Sims (1980). In contrast
time series procedures seem to be less operational in worlds where
policy interventions modify ()2s since it is easily shown that such
interventions will typically induce parameter instability in the
reduced form and, thereby, complicate the task of separating tran-
sient parameters from invariant ones (if any!). Simple theoretical
examples supporting this argumentation are found, e.g., in Engle
et al. (1983). Its empirical relevance within a given context is subject
to testing using the sort of data modern economies have managed
to provide us with!
interest rate short term coefficients. The interest rate impact coef-
ficient even changes sign across the two regimes. In contrast the
steady state long term solution of the demand for money equation
is invariant over the complete sample period though the policy
relevance of this empirical finding is limited since the equation
displays a slow adjustment speed towards its long term solution.
The evidence on the weak exogeneity of the interest rate variable
over the first regime is mostly inconclusive. In contrast, and some-
what counterintuitively, the weak exogeneity of interest rate over
the second regime cannot be rejected. (The weak exogeneity of
money over that regime had been rejected at an early stage of the
analysis). These conclusions are to be appraised in the light of the
difficulties encountered in the modelling of the interest equation but
it is worth noting that the OLS and FIML estimated coefficients
hardly differ from each other with the sole exception of the interest
rate impact coefficient.
(3) All together, it is clear that the predictive failure of the
demand for money equation across 1971 was not the result of an
invalid exogeneity assumption. The banking sector appears to have
modified substantially its behaviour in response to the introduction
of CCC in 1971 and one should account for the fact that it might
again do so if confronted to major changes in the monetary policy
of the Bank of England.
Appendix
M: Personal Sector M3
R: Local Authority 3 month deposit rate
Y: Personal disposable income
P: Implied deflator for personal disposable income
U: Unemployment rate
B: Real value of UK official reserves.
The data sources are given in Lubrano et al. (1986) wherefrom the
results which follow are taken.
337
[ - (0.12)
0.41 DI + 0.23 (1 - DI)] A I (1 R)
+ (0.09) L1 n + ,
- 0.32 DI~2 In (1 + R'_3) - 0.17 In (1 + R'_5)'
(0.09) (0.04)
338
References